References and source policy

Release candidate. The unit videos are added in the first release.

Source policy

This course teaches people to check claims. It would be indefensible if its own claims were unchecked. Three rules apply to everything written here.

1. Historical and mathematical claims carry a citation or an uncertainty label. Every factual claim in the course maps to an entry below or is plainly labelled contested or unsourced. A claim with neither is a defect that blocks release. Author-only drafting flags must be resolved before material reaches this reference list.

2. Claims about current AI carry a date. Model capabilities, tool names, pricing, institutional policy, benchmark results, and industry practice all age badly. Any such claim is written either so it does not date (a statement about how next-token prediction works) or with an explicit date attached (a statement about what a particular system did in a particular month). Undated claims about "what AI can do now" do not belong in the materials.

3. Named commercial systems are avoided unless pedagogically necessary. The course teaches a method that outlives any particular product. Naming a model version in the teaching text creates a maintenance burden and implies an endorsement. Where a learner needs an AI system for an activity, the material says "an AI assistant of your choice".

Verification status labels

Used throughout Historical Notes and the unit materials:

label meaning
established Supported by a primary source, or by consistent secondary sources with no serious dispute.
contested Sources disagree, or the claim is widely repeated but its evidential basis is unclear. State the disagreement; do not pick a side.
unsourced The source is not currently known. This does not mean false; state what evidence could settle it.

An author-only drafting flag is not a learner-facing uncertainty category and must not survive to release. Unsourced is the course's student-facing bucket, defined in Unit 16, for a claim deliberately left open together with the evidence that could move it.

Pre-release freshness review

Before any student-facing release, a reviewer must re-read every dated claim and either refresh it or remove it. It was done on 2026-10-31 for the v0.1.0 text, by re-opening each cited source; the access dates record it. It was done again on 2026-11-02 for the "as of" lines outside Unit 5, which rest on the provider documentation [D1]-[D4]: each page was re-opened. [D1], [D3] and [D4] were unchanged since their last revision (2026-10-23, 2026-10-30, 2026-10-07); [D2] had been revised that day and still says memory carries into new conversations and is on by default on the plans named. Every claim still held, so each was re-dated "as of November 2026" and nothing was corrected. Dates that record when one of the course's own checks was run (Units 15 and 17) are not "as of" claims and were left. If the release moves well past that date, do it again; the "as of November 2026" lines are the first to check.

Primary sources

[P1] Menabrea, L. F. (1842). "Notions sur la machine analytique de M. Charles Babbage." Bibliothèque Universelle de Genève, new series, vol. 41, no. 82, October 1842. The original French account of Babbage's Analytical Engine, based on Babbage's 1840 meetings in Turin. This is the article Lovelace translated. Accessed and read 2026-11-03: an offprint, not the journal volume (Accademia delle Scienze di Torino copy, https://archive.org/details/TO01056056, 27 pp., own pagination). Its title page says it is drawn from the Bibliotheque Universelle de Geneve, October 1842; offprint p. 4 says Babbage shared his ideas with Menabrea during a visit to Turin. It gives neither the year 1840 nor lectures; those rest on [P3, pp. 129-130], where Babbage dates the Turin meetings to 1840.

[P2] Menabrea, L. F., trans. A. A. L. [Augusta Ada Lovelace] (1843). "Sketch of the Analytical Engine Invented by Charles Babbage... with notes by the translator." In R. Taylor (ed.), Scientific Memoirs, vol. 3, London: Richard and John E. Taylor, pp. 666–731. The translation plus Lovelace's Notes A–G. Babbage described the Notes as "about three times the length of the original memoir" [P3, p. 136]; in this printing the memoir takes pp. 666 to 690 and the Notes pp. 691 to 731, 25 pages to 41. Note G contains the Bernoulli-number diagram and the "no pretensions whatever to originate any thing" passage (p. 722; "any thing" is two words as printed, where many later transcriptions write "anything" -- checked against the Internet Archive scan of vol. 3, india.history.resource.53369, on 2026-10-24). Public domain. The diagram itself is headed "for the computation by the Engine of the Numbers of Bernoulli. See Note G (page 722 et seq.)", so the Bernoulli material sits at page 722 and following; the surrounding Notes span the pp. 666–731 range commonly cited for the whole translation. Confirmed against the facsimile [P2a]. Opened 2026-11-03 (same scan, PDF pages 693-760): p. 666 gives the source as the Bibliotheque Universelle de Geneve, No. 82, October 1842; p. 669 says that in August 1840 Babbage had general plan No. 25 lithographed in Paris; pp. 676-678 describe the mill, the Jacquard cards and the Engine's operation cards and cards of the Variables, and p. 681 calls the columns of Variables "a store of numbers". Note A begins p. 691 and Note G p. 722, where both "originate any thing" and "whatever we know how to order it to perform" stand; pp. 725-727 label the numbers B1, B3, B5, B7 and set the diagram to compute B7 (n = 4) from B1, B3, B5. The diagram is an unnumbered folding plate between pp. 730 and 731.

[P3] Babbage, C. (1864). Passages from the Life of a Philosopher. London: Longman, Green, Longman, Roberts, & Green. Babbage's own retrospective account, including his description of the Analytical Engine and of Lovelace's work on the Notes. A primary source that is also an interested party — useful, not neutral. Accessed and read 2026-11-03: the 1864 London edition scanned at https://archive.org/details/passagesfromlife00babb, read with the Project Gutenberg transcription (ebook 57532). Ch. VIII, "Of the Analytical Engine": p. 117 says the Engine "consists of two parts", the store and the mill, and pp. 117-118 describe its operation cards and variable cards; pp. 129-130 date the Turin meetings to 1840; p. 136 has the Lovelace passage ("having detected a grave mistake which I had made in the process"; "about three times the length of the original memoir"). P. 449 (1864): "If I survive some few years longer, the Analytical Engine will exist".

Facsimiles

The Note G table must be checked against a facsimile rather than a transcription, because the error the course discusses is precisely the kind of thing a transcription silently corrects.

[P2a] Facsimile of the Note G diagram. Photograph of the printed "Diagram for the computation by the Engine of the Numbers of Bernoulli" from Scientific Memoirs vol. 3, as offered by Sophia Rare Books: https://www.sophiararebooks.com/pictures/3544a.jpg (accessed 2026-07-23; 1000×654 JPEG). It was retained at site/assets/images/facsimile-noteg.jpg, so that the check did not depend on the URL persisting, until 2026-10-31, when the copy below replaced it there. Legible from that photograph: the diagram heading and page reference (722 et seq.); that the error is in operation 4; and that operation 4 acts on ²V₅ and ²V₄ in that order. Not resolvable at this resolution: the fine typography of the division operator, which is easy to misread. For anything turning on the exact glyph, a higher-resolution facsimile or the physical volume is needed. See Historical Notes §4.

Since 2026-10-31 the course shows a different copy of the diagram: Wikimedia Commons, "File:Diagram for the computation of Bernoulli numbers.jpg", https://commons.wikimedia.org/wiki/File:Diagram_for_the_computation_of_Bernoulli_numbers.jpg (downloaded 2026-10-31; 2000x1400 JPEG, retained at site/assets/images/facsimile-noteg.jpg). It is a processed image of the printed plate, not an unaltered photograph. Commons tags it Public domain, names Ada Lovelace as the artist, and credits it to http://www.sophiararebooks.com/pictures/3544a.jpg, the address of the photograph above; its relation to that photograph is only what this credit states. The readings above were made from the earlier photograph. Legible in this copy (read 2026-10-31): the division operator of operation 4, a division sign; and row 24 as "+ 4V13 + 0V24 -> 1V24", stated as "= B7", with no change of sign. Re-read 2026-11-03: the same heading, operation 4 and row 24 as above. The plate is one unnumbered page, so it is cited without a locator; its rows are numbered by operation.

Higher-resolution scanned copies of Scientific Memoirs vol. 3 are also available through several digital library collections and should be recorded here if used.

[P4] Babbage, C. (1840). "The General Plan of Mr. Babbages Great Calculating Engine", Plan 25, dated 6 August 1840. Babbage's general plan of the Analytical Engine, shown in Unit 2 from a photograph: ArnoldReinhold, "File:Babbage Analytical Engine Plan 1840 CHM.agr.jpg", Wikimedia Commons, CC BY 4.0, https://commons.wikimedia.org/wiki/File:Babbage_Analytical_Engine_Plan_1840_CHM.agr.jpg (accessed and downloaded 2026-11-02; 3796x2557 JPEG, resized to 2000x1347, re-encoded and retained at site/assets/images/babbage-plan-25-1840.jpg). The file page describes it as "Photo of Babbage Analytical Engine Plan from 1840 at the Computer History Museum", dates it 6 August 1840, and gives the source as the photographer's own work. The title and plan number are read from the photograph (2026-11-02): the cartouche "The General Plan of Mr. Babbages Great Calculating Engine." and, at the centre, "Plan 25. 6 August 1840." Whether the museum shows the original drawing or a copy of it the file page does not say. Opened 2026-11-03: the file page and the retained photograph are unchanged on these points. [P2, p. 669] independently records that in August 1840 Babbage had "one of his general plans (No. 25)" of the Engine lithographed in Paris.

Secondary sources

[S1] to [S5] have not been opened by this course: no open copy was found on 2026-11-03 ([S1] is behind the publisher's paywall; [S2] to [S5] are books). The register claims once resting on them alone are located in sources that were opened: H1 in [S12] and [S7]; H2 in [P2] and [P3]; H9 and H10 in [S10] and [S11], which report the positions of [S2], [S3] and [S4] with page references; H18 in [S7]. [S1] to [S5] remain cited as the works those positions come from.

[S1] Bromley, A. G. (1982). "Charles Babbage's Analytical Engine, 1838." Annals of the History of Computing, 4(3), 196–217. Technical reconstruction of the Engine's architecture. The standard reference for what the machine could actually do.

[S2] Stein, D. (1985). Ada: A Life and a Legacy. Cambridge, MA: MIT Press. A sceptical assessment of Lovelace's mathematical contribution. Represents one pole of a genuine and continuing scholarly disagreement.

[S3] Toole, B. A. (1992). Ada, the Enchantress of Numbers: A Selection from the Letters of Lord Byron's Daughter and Her Description of the First Computer. Mill Valley, CA: Strawberry Press. A collection of correspondence, taking a considerably more expansive view of Lovelace's contribution than [S2]. Read alongside it, not instead of it.

[S4] Swade, D. (2000). The Difference Engine: Charles Babbage and the Quest to Build the First Computer. New York: Viking. On Babbage's machines and the modern reconstruction effort.

[S5] Hollings, C., Martin, U., and Rice, A. (2018). Ada Lovelace: The Making of a Computer Scientist. Oxford: Bodleian Library. A recent treatment drawing on the Lovelace archives, notable for engaging directly with the disputed questions rather than taking a side by omission. Useful for Unit 16's distinction between established and contested claims.

[S6] The printed errors, and what each reading computes. Operation 4 is printed as V₅ ÷ V₄ where V₄ ÷ V₅ is correct, read by [S7, sec. 5.2] as a typesetting rather than an authorial error; its location and operand inversion are established by the facsimile [P2a] and corroborated by the reconstruction [S7]. It is not the only printed fault: [S7] §5.2 also lists operation 24, where "the negative value of V₁₃ must be transferred to V₂₄", and defects in operation 21. A phrase like "the table as printed" therefore does not name one object.

The course reconstructed the 25-operation table and ran it with exact arithmetic (course/lab/noteg.py), which reproduces the documented −1/30 when both faults are repaired and then reports each other reading: 139/630 (operation 4 as printed, 24 repaired, earlier values supplied), −139/630 (both faults literal), 1/30 (only 24 as printed), and −25621/630 when the routine, with operation 4 as printed and operation 24 repaired, generates the earlier Bernoulli numbers it consumes.

The last of those is the value repeated in encyclopaedia and blog sources (https://en.wikipedia.org/wiki/Note_G, accessed 2026-07-23). Correction of record: this course previously stated that 139/630 was what the as-printed table computes and that −25621/630 "does not survive recomputation". Both were wrong for the same reason — neither said which object was executed — and the 139/630 run silently repaired operation 24. What those sources can still be faulted for is not stating their interpretation; so could this course, until 2026-07-26. Re-opened 2026-11-03: the article (revision of 2026-08-15), section "Error", still gives -25621/630 without saying which reading it runs.

[S7] Glaschick, R. (2016). Ada Lovelace's Calculation of Bernoulli's Numbers. Paderborn, 29 September 2016. https://rclab.de/_media/analyticalengine/aal_noteg_glaschick_v1.2.pdf (accessed 2026-07-23). A detailed independent reconstruction, self-published — a step-by-step working of the Note G table that states the operation-4 operand swap directly ("V₄ must be divided by V₅") and groups it with typographical-domain errors; independently confirms the target value B₇ = −1/30. Opened 2026-11-03 (18-page print of the web page, cited by section): sec. 2 says "no Analytical Engine was ever built"; sec. 3.3 gives the 25 operations with the corrections to lines 4 and 24; sec. 5.1 gives Lovelace's B1, B3, B5, B7 as 1/6, -1/30, 1/42, -1/30; sec. 5.2 lists the errors under "a few minor errors that are in the domain of typographical errors"; sec. 5.3 works B7 = -1/30 by hand. Supports: the operation-4 swap, the operation-24 and operation-21 faults, and that the operation-4 error is classed as typographical (H18). May NOT say: who introduced it. Sec. 5.2 ends by calling the errors glitches that "do not disqualify Ada Lovelace", which treats them as hers rather than the printer's.

[S8] Campbell-Kelly, M. (ed.) (1989). The Works of Charles Babbage, vol. 3. London: Pickering. Cited by [S7] (p. 1) as the fullest correction of the Note G table (p. 159). Not consulted by this course: nothing here rests on it alone.

[S9] O'Connor, J. J., and Robertson, E. F. (1998). "Jacob (Jacques) Bernoulli." MacTutor History of Mathematics, School of Mathematics and Statistics, University of St Andrews. Last update September 1998. https://mathshistory.st-andrews.ac.uk/Biographies/Bernoulli_Jacob/ (accessed 2026-11-02). A reference biography, opened and read in full. It says Ars Conjectandi was published in Basel in 1713, eight years after his death, and "The Bernoulli numbers appear in the book". Supports: the numbers were in print by 1713, well before Note G (1843). May NOT say: who first found them (MacTutor's biography of Seki says he reached them before Bernoulli), how Bernoulli derived them, or that the 1713 book used the modern index or either convention. Re-opened 2026-11-03: unchanged; both statements are in the paragraph on Ars Conjectandi. The page has no headings or numbered paragraphs to cite.

[S10] Hollings, C., Martin, U., and Rice, A. (2017). "The early mathematical education of Ada Lovelace." BSHM Bulletin: Journal of the British Society for the History of Mathematics, published online 1 June 2017, doi 10.1080/17498430.2017.1325297. Open access (CC BY-NC-ND). Accessed and read 2026-11-03 as the online-first PDF hosted by the Clay Mathematics Institute, https://www.claymath.org/wp-content/uploads/2023/04/the_early_mathematical_education_of_ada_lovelace.pdf; its pages are not those of the printed volume, so it is cited by section. The abstract describes the Engine as "unbuilt" and Note G's table as "often described as 'the first computer program'". Section "Changing perceptions of Lovelace" reports Stein (1985, p. 90) arguing that her command of mathematics was limited, Swade (2000, p. 166) calling her a "talented beginner, a precocious novice", Toole (1992, p. 2) presenting her as "a synthesizer and a visionary", Bromley (1990) finding no evidence she prepared a program, and Misa (2016) concluding the Notes were "a product of an intense intellectual collaboration". Supports: that the extent of her contribution is disputed, and by whom (H10), and that "first program" is a common description (H9). May NOT say: how the work was in fact divided, or that Note G is or is not the first program; the authors take a side, against Stein and Swade, on her mathematical ability.

[S11] Hollings, C., Martin, U., and Rice, A. (2017). "The Lovelace-De Morgan mathematical correspondence: A critical re-appraisal." Historia Mathematica, doi 10.1016/j.hm.2017.04.001. Open access (CC BY). Accessed and read 2026-11-03 as the article-in-press PDF hosted by the Clay Mathematics Institute, https://www.claymath.org/wp-content/uploads/2023/04/Lovelace-De-Morgan-correspondence.pdf, paginated 1-30 rather than as printed, so cited by section. The abstract says the Engine, "had it been built, would have been a general-purpose computer", and that there has been "considerable disagreement among scholars as to her mathematical proficiency". Sec. 1 says the table is "often called the 'first computer program'" (note 2 adds that in modern terms it is an execution trace, the program being the stack of cards), and sets out the debate: Stein (1985), Bromley (1982, 1990), Misa (2016), Toole (1992). Sec. 5 corrects the dating of letters as transcribed by Toole and Stein. Supports: H9 and H10 as for [S10]. May NOT say: who wrote which part of the Notes. Sec. 6 takes the paper to be a collaboration with Babbage as a working view, and argues against Stein that Lovelace had the skill to contribute; that is the authors' position, not a settled fact.

[S12] Science Museum Group Collection. "Babbage's Analytical Engine, 1834-1871. (Trial model)", object number 1878-3. https://collection.sciencemuseumgroup.org.uk/objects/co62245/babbages-analytical-engine-1834-1871-trial-model-analytical-engine. Accessed 2026-11-03 through a fetch tool that returns a summary with short quotations; a direct download was refused (HTTP 403), so the page was not read raw. As returned: "Only part of the machine as a trial piece was completed before Babbage's death in 1871"; the object is a portion of the mill with printing mechanism, under construction when he died; the machine was conceived in 1834. Supports: the Engine was never completed in Babbage's lifetime (H1). May NOT say: anything about its architecture beyond the mill, or about Note G.

A note on tertiary sources

Popular articles, encyclopaedia entries, and blog posts about Note G's table error are abundant, mutually inconsistent, and frequently trace back to each other rather than to the table. They are useful for finding out what is commonly claimed — which Unit 16 asks learners to examine — and are not sufficient support for a factual assertion in the materials. Where the course reports a common claim, it says that it is a common claim.

Mathematical references

[M1] Bernoulli numbers, conventions. Two standard conventions differ in the value of B1: the first Bernoulli numbers (B_n^-) take B1 = -1/2, and the second (B_n^+) take B1 = +1/2. All other values agree. This course fixes the first convention; see course/lab/reference_bernoulli.py. The first are generated by t/(e^t - 1) (DLMF eq. 24.2.1, see [M6]); the second by t/(1 - e^(-t)), which equals t/(e^t - 1) + t and so changes only the t coefficient, B1. That identity is the course's derivation, not a DLMF statement; the two series were expanded with exact fractions through B11 on 2026-11-02 and agree with the oracle except at B1. Re-opened 2026-11-03: DLMF eq. 24.2.1 still gives t/(e^t - 1) = sum_{n>=0} B_n t^n/n! for |t| < 2 \pi.

[M2] Published value tables. The oracle's PUBLISHED_TABLE was checked against NIST Digital Library of Mathematical Functions, Release 1.2.7 of 2026-06-15, F. W. J. Olver et al. (eds.), Chapter 24, §24.2(iv), Tables 24.2.1 and 24.2.3, National Institute of Standards and Technology, https://dlmf.nist.gov/24.2.T1 and https://dlmf.nist.gov/24.2.T3 (accessed 2026-07-24). The subsection identifies its printed source as M. Abramowitz and I. A. Stegun (eds.) (1964), Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables, National Bureau of Standards Applied Mathematics Series 55, Washington, DC: U.S. Government Printing Office, pp. 809–810; the electronic source corresponds to the corrected tenth printing of December 1972.

The comparison covers B0–B12, every even value through B30 = 8615841276005/14322, and the odd zeros represented in the oracle. The table and DLMF equation 24.2.2 use the course's convention B1 = -1/2. All 22 stored values agree exactly; no correction was needed. Re-opened 2026-11-03 (Version 1.2.8, 2026-09-15): the sec. 24.2(iv) note still names Abramowitz and Stegun (1964, pp. 809-810), and Tables 24.2.1 and 24.2.3 still agree with all 22 values. The tenth-printing statement above was not re-checked.

[M3] Defining recurrence. The oracle uses sum_{j=0}^{n} C(n+1, j) * B_j = 0 for n >= 1, with B0 = 1, which yields B1 = -1/2 directly. course/lab/examples/correct_example.py deliberately uses a different method (the Akiyama–Tanigawa transform) so that agreement between them is evidence rather than duplication.

Language models and current AI systems

These references support the compact account in Unit 3. They describe distinct layers of a modern assistant: transformer architecture, pre-training and post-training, retrieval, and calibrated self-evaluation. Unit 3 does not imply that every product uses every technique.

[L1] Vaswani, A. et al. (2017). "Attention Is All You Need." Advances in Neural Information Processing Systems, 30. arXiv 1706.03762. https://arxiv.org/abs/1706.03762 (accessed 2026-07-24). Introduced the Transformer architecture. Its attention mechanism combines information from different token positions to produce context-dependent representations. This citation supports the architecture sketch in Unit 3, not a claim that attention provides a truth check or a human-readable explanation. Opened 2026-11-03 (arXiv v7): sec. 3 introduces the Transformer; sec. 3.2 defines attention as a weighted sum of values, and sec. 3.2.3 lets each decoder position attend only to positions up to and including itself.

[L2] Ouyang, L. et al. (2022). "Training Language Models to Follow Instructions with Human Feedback." Advances in Neural Information Processing Systems, 35, 27730–27744. arXiv 2203.02155. https://arxiv.org/abs/2203.02155 (accessed 2026-07-24). Shows one influential post-training pipeline: supervised demonstrations followed by preference data and reinforcement learning from human feedback. It supports the distinction between next-token pre-training and training an assistant to follow instructions. Opened 2026-11-03 (arXiv v1): sec. 3.1 sets out labeler demonstrations for supervised fine-tuning, comparisons of model outputs for a reward model, and PPO; sec. 3.5 has labelers rank 4 to 9 responses per prompt.

[L3] Lewis, P. et al. (2020). "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks." Advances in Neural Information Processing Systems, 33, 9459–9474. arXiv 2005.11401. https://arxiv.org/abs/2005.11401 (accessed 2026-07-24). Combines a pre-trained generator with an external document index. It supports Unit 3's distinction between generation from model parameters and a system that also retrieves inspectable sources. Opened 2026-11-03 (arXiv v4): sec. 2 describes a retriever that returns text passages for the input and a generator conditioned on the input and a retrieved passage; the abstract names the index as Wikipedia.

[L4] Kadavath, S. et al. (2022). "Language Models (Mostly) Know What They Know." arXiv 2207.05221. https://arxiv.org/abs/2207.05221 (accessed 2026-07-24). Finds that specially elicited or trained self-evaluation scores can discriminate correct from incorrect answers on some evaluated tasks, with calibration and out-of-distribution limits. This is why Unit 17 says ordinary confident wording is not a correctness defence while avoiding the stronger, false claim that every model-derived confidence measure carries no information. Opened 2026-11-03 (arXiv v4): sec. 1.1 summarises that calibration improves with model size and few-shot prompts, depends on answer format, needs a temperature adjustment for RLHF policies, and is poor for P(IK) on new distributions. The paper measures model probabilities (P(True), P(IK)), not confident wording.

Evaluation and experimental method

Background for the evaluation segment in Unit 8 and for Evaluating, when one check is not enough, which is the overview behind it. The course teaches the protocol and points at these; it does not teach the methods themselves. The entries are here so that the overview's claims about what exists can be checked, and so that a reader who is going to run a real evaluation knows which literature to enter.

[E1] Strathern, M. (1997). "'Improving ratings': audit in the British University system." European Review, 5(3), 305–321. The source of the widely quoted formulation of Goodhart's law — that a measure which becomes a target ceases to be a good measure. Cited for the formulation, not for the audit-culture argument around it. Not opened 2026-11-03: the publisher's text is paywalled and no repository copy was found, so the page of the formulation is not confirmed.

[E2] Nosek, B. A., Ebersole, C. R., DeHaven, A. C., and Mellor, D. T. (2018). "The preregistration revolution." Proceedings of the National Academy of Sciences, 115(11), 2600–2606. What preregistration is, what it does and does not fix, and the distinction between prediction and postdiction that the freeze step rests on. Opened 2026-11-03: the published abstract (p. 2600, via PubMed) defines preregistration as defining the research questions and analysis plan before observing the outcomes; in the OSF preprint, "Preregistration distinguishes prediction and postdiction" adds that the plan is usually posted to an independent registry.

[E3] Cumming, G. (2014). "The New Statistics: Why and How." Psychological Science, 25(1), 7–29. The case for estimation — effect sizes and intervals — over significance testing, and a practical account of how to report results that way. Opened 2026-11-03: the published abstract only (p. 7, via PubMed), which calls for a shift from NHST to estimation based on effect sizes, confidence intervals and meta-analysis; the paywalled body was not opened.

[E4] Button, K. S., et al. (2013). "Power failure: why small sample size undermines the reliability of neuroscience." Nature Reviews Neuroscience, 14, 365–376. Why an underpowered study that finds an effect has probably overestimated it. The argument is general, not specific to neuroscience. Opened 2026-11-03: the published abstract only (p. 365, via PubMed): low power lowers the chance that a significant result reflects a true effect, and its consequences include overestimated effect sizes and low reproducibility; the paywalled body was not opened.

[E5] Ioannidis, J. P. A. (2005). "Why Most Published Research Findings Are False." PLoS Medicine, 2(8), e124. The structural argument: small samples, small effects, many tested relationships and flexibility in design together make a positive finding unreliable. Opened 2026-11-03 (open-access PDF): Corollaries 1 and 2 (p. 697) and 3 and 4 (p. 698) give the small-study, small-effect, many-relationships and flexibility arguments; the paper does not discuss inflated effect sizes.

[E6] Kapoor, S. and Narayanan, A. (2023). "Leakage and the reproducibility crisis in machine-learning-based science." Patterns, 4(9), 100804. A taxonomy of the ways test data contaminates training in practice, with a survey of affected published work. The reference for the held-out-data warning. Opened 2026-11-03 (Patterns full text via Europe PMC): "Toward a solution: A taxonomy of data leakage" lists eight types, among them pre-processing on training and test set (L1.2) and duplicates (L1.4); "Data leakage causes irreproducible results" reports 294 affected papers in 17 fields.

[E7] Krippendorff, K. (2018). Content Analysis: An Introduction to Its Methodology, 4th edition. Thousand Oaks, CA: Sage. The standard treatment of building a coding scheme and of reliability, including Krippendorff's α. Not opened 2026-11-03: a book with no open copy, so no chapter or page is confirmed.

[E8] Cohen, J. (1960). "A Coefficient of Agreement for Nominal Scales." Educational and Psychological Measurement, 20(1), 37–46. Cohen's κ: agreement between two raters, corrected for the agreement chance alone would produce. Not opened 2026-11-03: paywalled, and no repository copy was found.

[E9] Braun, V. and Clarke, V. (2006). "Using thematic analysis in psychology." Qualitative Research in Psychology, 3(2), 77–101. A step-by-step account of thematic analysis, and the standard entry point to qualitative coding for people who have not done it before. Not opened 2026-11-03: the publisher's text is paywalled and the UWE repository copy refused automated access (HTTP 403), so nothing cited to it is confirmed.

[E10] Myers, G. J., Sandler, C., and Badgett, T. (2011). The Art of Software Testing, 3rd edition. Hoboken, NJ: Wiley. The classic general text. Its opening argument — that a successful test is one that finds an error — is the disposition the whole of this page assumes. Not opened 2026-11-03: a book with no open copy, so the chapter is not confirmed.

[E11] Claessen, K. and Hughes, J. (2000). "QuickCheck: a lightweight tool for random testing of Haskell programs." Proceedings of the Fifth ACM SIGPLAN International Conference on Functional Programming (ICFP '00), 268–279. The paper that established property-based testing. The idea, not the language, is what matters here; implementations exist for most languages. Not opened 2026-11-03: the ACM copy refused automated access and no author-hosted copy was found.

[E12] McKeeman, W. M. (1998). "Differential Testing for Software." Digital Technical Journal, 10(1), 100–107. Comparing independent implementations on the same inputs. This is what the course's own check.py does; Unit 8's first-divergence report is the course's addition, not McKeeman's (he localises a fault by reducing the test, p. 105). Opened 2026-11-03 (archived HP copy of the journal PDF): p. 101 defines differential testing as feeding one test to several comparable programs and treating a differing result as a possible bug; p. 105 localises the bug by reducing the test to the shortest version that still shows it. It does not describe a first-divergence report.

[E13] Chen, T. Y., Cheung, S. C., and Yiu, S. M. (1998). Metamorphic Testing: A New Approach for Generating Next Test Cases. Technical report HKUST-CS98-01, Department of Computer Science, Hong Kong University of Science and Technology. Testing when no oracle is available, by asserting relations between outputs rather than the value of any one output. It applies to generative systems, where a correct answer usually cannot be written down. Opened 2026-11-03 via arXiv 2002.12543, which reproduces the report: secs. 1 and 4 present the method as usable without a test oracle; sec. 3.3 (shortest path) derives follow-up test cases whose expected results are relations to the first output, such as two distances that must sum to it.

[E14] Jia, Y. and Harman, M. (2011). "An Analysis and Survey of the Development of Mutation Testing." IEEE Transactions on Software Engineering, 37(5), 649–678. Deliberately introducing faults to measure whether a test suite would catch them. The survey rather than the original papers, because it is the readable entry point. Not opened 2026-11-03: paywalled, and no authorised open copy was found.

[E15] Zheng, L., et al. (2023). "Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena." Advances in Neural Information Processing Systems 36 (NeurIPS 2023), Datasets and Benchmarks Track. Measures both the agreement between model judgements and human ones and the biases in model judgement — position, verbosity and self-preference. Cited with its date, because it is a claim about particular systems at a particular time. Opened 2026-11-03 (arXiv v4): sec. 3.3 reports position bias (most judges favour the first position) and verbosity bias, and says its data cannot determine whether self-enhancement bias exists, though GPT-4 and Claude-v1 favoured their own answers.

[E16] Shadish, W. R., Cook, T. D., and Campbell, D. T. (2002). Experimental and Quasi-Experimental Designs for Generalized Causal Inference. Boston: Houghton Mifflin. The standard treatment of design and of threats to validity. Where to start if your evaluation involves people rather than only artefacts. Not opened 2026-11-03: a book with no open copy.

Second-order reasoning

Background for Second-Order Reasoning: The Lens Behind The Course and Unit 18. Read that page first: it marks the boundary between the established results below and the course's own synthesis, and the course's claims discipline applies to its own capstone exactly as it applies to everything else.

[T2] Tarski, A. (1936). "Der Wahrheitsbegriff in den formalisierten Sprachen." Studia Philosophica, 1, 261–405. The undefinability of truth: under the theorem's formal conditions, arithmetical truth for the object language is not definable within that same language. This is a scoped result about formal languages. Applying it to a language model, a person, or an organisation requires a separate argument and is not a corollary of Tarski's theorem. Not opened 2026-11-03: no openly available copy of any of the three texts below was found (the Polish original at the Kujawsko-Pomorska Biblioteka Cyfrowa is restricted to the UMK campus; the Internet Archive copy of Logic, Semantics, Metamathematics, 1956, is lending-only), so no page is given for the theorem.

Publication history, checked. Volume and page range are established: the German text is Studia Philosophica vol. 1, pp. 261–405, Leopold Blaustein's translation of the Polish original with a postscript added (Stanford Encyclopedia of Philosophy, "Alfred Tarski", bibliography, https://plato.stanford.edu/entries/tarski/, accessed 2026-07-23; that bibliography lists the German text as 1935, which is the datum the contested-year discussion below turns on). The year is contested, and the disagreement is recorded here rather than settled. The volume is dated 1935 and the Stanford Encyclopedia of Philosophy cites it as 1935; Peter Milne, "On the Year of Publication of Tarski's 'Der Wahrheitsbegriff in den formalisierten Sprachen'", History and Philosophy of Logic, 46(2), 273–286 (DOI 10.1080/01445340.2024.2334173, published online 26 April 2024), argues from correspondence that the journal issue did not appear until 1936, and that what circulated in late 1935 were preprints lacking two corrections and a short addendum. The course cites 1936; a reader who finds 1935 elsewhere has not found an error.

The other versions, and why a page number depends on which one. The Polish original is Pojęcie prawdy w językach nauk dedukcyjnych, Warsaw: Nakładem Towarzystwa Naukowego Warszawskiego, 1933, VII + 116 pp., in the series Prace Towarzystwa Naukowego Warszawskiego, Wydział III: Nauk Matematyczno-Fizycznych, no. 34 (catalogue record, Kujawsko-Pomorska Biblioteka Cyfrowa, permalink https://kpbc.umk.pl/dlibra/publication/265521/edition/262893, accessed 2026-07-23, which confirms the series title, the Warsaw imprint and the VII + 116 pagination; the "Wydział III" number 34 as cited by the Stanford Encyclopedia of Philosophy). The English translation, "The Concept of Truth in Formalized Languages", is in Logic, Semantics, Metamathematics, 2nd edn 1983 [1956], pp. 152–278. Three texts, three paginations: a page reference to "Tarski" means nothing unless it says which one.

[T3] Gödel, K. (1931). "Über formal unentscheidbare Sätze der Principia Mathematica und verwandter Systeme I." Monatshefte für Mathematik und Physik, 38, 173–198. The first incompleteness theorem. Closely related to [T2]; the two are frequently conflated in secondary accounts, and Unit 18 should keep them distinct. The same paper also states the second incompleteness theorem, that such a system cannot prove its own consistency: the final version Gödel sent for publication "now contained also a statement of the second incompleteness theorem" ([T24], sec. 5). Opened 2026-11-03, a page scan of the 1931 printing (Internet Archive item uber-formal-unentscheidbare-satze-der-principia-mathematica-und-verwandter-systeme-i_20250205): Satz VI, p. 187, gives undecidable sentences for every omega-consistent recursive class of formulas; p. 189 shows that the unprovability half uses only consistency, which p. 196 restates; pp. 190-191 say the proof uses only recursive axioms and rules and the definability of every recursive relation, and name the Peano axiom system among the systems that qualify; Satz XI, p. 196, is the second theorem. Rosser's p. 87 [C8] places Satz VI and Satz XI on the same pages.

[T7] von Foerster, H. (2003). Understanding Understanding: Essays on Cybernetics and Cognition. New York: Springer-Verlag. Collected essays from the tradition that includes the observer in the system being described. Confirmed against the book's own front matter: "© 2003 Springer-Verlag New York, Inc.", ISBN 0-387-95392-2, xii + 362 pp., with the printing line "9 8 7 6 5 4 3 2 1" — the first printing, and there is no later edition. The Library of Congress CIP data on the same page carries the year 2002 (Q315.5 .V64 2002), which is why some catalogues give 2002; the imprint year is 2003. Chapter and page references used by the course are to this printing. The 2003 Springer edition is also online with per-chapter DOIs under the stem 10.1007/0-387-21722-3 (e.g. chapter 14, "Ethics and Second-Order Cybernetics", pp. 287–304, DOI 10.1007/0-387-21722-3_14; the chapter DOI, its title, page range and 2003 year confirmed via the Crossref record, accessed 2026-07-23).

[T8] Maturana, H. R. and Varela, F. J. (1980). Autopoiesis and Cognition: The Realization of the Living. Boston Studies in the Philosophy of Science, vol. 42. Dordrecht, Holland: D. Reidel. The companion reference for the same tradition. Confirmed against the book's own copyright page: published by D. Reidel Publishing Company, P.O. Box 17, Dordrecht, Holland; series Boston Studies in the Philosophy of Science, vol. 42; copyright 1980; ISBN 90-277-1015-5 (cloth) and 90-277-1016-3 (paperback). The same 1980 volume is reprinted by Springer under the stable DOI 10.1007/978-94-009-8947-4, whose Crossref record (accessed 2026-07-23) confirms the title, the 1980 date, ISBN 978-90-277-1016-1 and the Boston Studies series — the Crossref record uses the later series name "Boston Studies in the Philosophy and History of Science", the same renaming noted below. No edition statement is printed — this is the first English book publication. Sold and distributed in the USA and Canada by Kluwer Boston Inc., Hingham, MA, which is why several catalogues record the place of publication as Boston; the imprint is Dordrecht. The series was later renamed Boston Studies in the Philosophy and History of Science, and records that use the newer name for volume 42 are describing the same book. The second of the two essays, "Autopoiesis: The Organization of the Living" (from p. 63), first appeared in Spanish as De Máquinas y Seres Vivos, Editorial Universitaria S.A., 1972.

[T9] Madaan, A. et al. (2023). "Self-Refine: Iterative Refinement with Self-Feedback." Advances in Neural Information Processing Systems, 36. Experiments on seven tasks found that iterative feedback and revision using the same language model could improve outputs over one-step generation. This is dated, task-bounded evidence that same-model revision can be useful; it does not make the revision independent verification. Primary paper: https://proceedings.neurips.cc/paper_files/paper/2023/hash/91edff07232fb1b55a505a9e9f6c0ff3-Abstract-Conference.html (accessed 2026-11-03). Opened 2026-11-03, the proceedings PDF: the abstract and sec. 3 give the seven tasks and the gain over one-step generation; sec. 3.3 reports only modest gains on math reasoning because the model could not tell whether its own answer was wrong (ChatGPT's feedback was "everything looks good" for 94% of instances), and larger gains when an external source flagged the error.

A note on how these are used. [T2] and [T3] are established results; they are stated as established. [T3]'s citation is settled; [T2]'s year is contested, as its entry records. [T7] and [T8] are named as a research tradition rather than as results. [T9] is dated empirical evidence, not a general guarantee about self-correction. Unit 18's checkpoint asks the learner to sort these categories themselves, so the categories must be visible in the material rather than smoothed into a single voice.

Second-order cybernetics

Background for Second-Order Reasoning: The Lens Behind The Course and Unit 18, which cite these as [C1] to [C9]. They are the sources for the cybernetic tradition itself, as distinct from the formal results at [T2] and [T3]. All of it is established, published material, and none of it is a claim about language models.

[C1] von Foerster, H. — second-order cybernetics, the cybernetics of observing systems. The distinction between the cybernetics of observed systems, where the observer stands outside, and the cybernetics of observing systems, where the observer is included in what is described, is formulated in his work. The usual English source is the collected papers Understanding Understanding: Essays on Cybernetics and Cognition, New York: Springer, 2003 — the same volume listed above as [T7]. Cite [C1] for the first-order/second-order distinction and [T7] for the volume; they are one source under two uses.

The formulation, located. It is stated in von Foerster's own words in "Ethics and Second-Order Cybernetics", chapter 14 of [T7] (DOI 10.1007/0-387-21722-3_14, Crossref record accessed 2026-07-23), pp. 287–304 — first published in French in Systèmes, Éthique, Perspectives en thérapie familiale, ed. Y. Rey and B. Prieur, Paris: ESF éditeur, 1991, pp. 41–55, per the chapter's own source note. Opened 2026-11-03: Springer's free preview shows pp. 287-288 only; there the chapter opens as a conference address ("Ladies and Gentlemen", "opening the Plenary sessions"), not an interview, and the source note prints the editors as "Y. Ray et B. Prieur". Pp. 289-304 were not openly available, so the wording below was not re-checked against pp. 299 and 303. On p. 303: "I would say, first-order cybernetics is the cybernetics of observed systems, while second-order cybernetics is the cybernetics of observing systems." The first half of the pair is given separately on p. 299: "the whole conceptual machinery of 'early' cybernetics, first-order cybernetics; or as I would say, the cybernetics of observed systems." Cite one of those two pages; do not attribute the formulation to the volume at large. (Chapter 13 of the same volume, pp. 283–286, is von Foerster's own "Cybernetics of Cybernetics", which is a different text from Mead's at [C2].)

The Biological Computer Laboratory. Much of the work was done there. It was founded by von Foerster at the University of Illinois at Urbana-Champaign in 1958 and closed in 1976 — "founded by Professor Heinz von Foerster in 1958, made Illinois an international center of cybernetics research until the lab's demise in 1976" (Illinois Distributed Museum, University of Illinois, https://distributedmuseum.illinois.edu/exhibit/biological_computer_laboratory/, accessed 2026-07-23). The 1958 founding is confirmed independently by the University of Illinois Archives: "he established the Biological Computer Laboratory in 1958" (https://archives.library.illinois.edu/2016/12/06/heinz-von-foerster-and-the-bcl/, accessed 2026-07-23). Give 1958–1976 rather than "the late 1950s to the mid 1970s".

[C9] Wiener, N. (1948). Cybernetics: Or Control and Communication in the Animal and the Machine. Cambridge, MA: The Technology Press; New York: John Wiley & Sons; Paris: Hermann. Opened 2026-11-03: the Open Library work record OL4307531W (this title, first published 1948; these publishers among its editions). The publisher's page refused automated access and the book itself was not read. Supports: the title and subtitle, and the year, as the course's one-line definition of the field. May NOT say: who coined the word, or when (the sources differ: Wiener alone in 1948, or with Rosenblueth in 1947); anything about the book's content beyond its title.

[C2] Mead, M. (1968). "Cybernetics of Cybernetics." In H. von Foerster, J. D. White, L. J. Peterson and J. K. Russell (eds), Purposive Systems: Proceedings of the First Annual Symposium of the American Society for Cybernetics, New York: Spartan Books, pp. 1–11. Commonly cited as the point at which the reflexive turn — cybernetics applied to itself — was named. The address is associated with the founding of the American Society for Cybernetics. The course states only that it is commonly cited as that point, which is what the available sources support. Opened 2026-11-03: the society's page prints only the closing excerpt of the paper, without page numbers, in which Mead proposes that the new society apply its own ideas to itself ("to ITSELF"); no page is given for that reason. The venue, editors and year are given in this form by the American Society for Cybernetics itself (https://asc-cybernetics.org/CofC/wp-content/uploads/2010/08/mead.html, accessed 2026-07-23, which prints the citation "Mead, M (1968) Cybernetics of Cybernetics, in Foerster, H von, White, J, Peterson, L and Russell, J (eds) (1968) Purposive Systems, New York, Spartan Books"); the page range is given by P. Pangaro, "Why do we want to live in cybernetics?", Cybernetics and Human Knowing, 23(2), 2016, 9–21, whose reference list reads "Mead, M. (1968). Cybernetics of cybernetics. In von Foerster, H., White, J. D., Peterson, L. J., & Russell, J. K. (Eds.), Purposive Systems (pp. 1–11). New York: Spartan Books", and who quotes Mead at pp. 9 and 10 — in-text page citations that fall inside the stated range. Contested: the occasion of the address. The society's first annual symposium is usually dated 26–27 October 1967 at the National Bureau of Standards, Gaithersburg, Maryland, but descriptions of the volume also place the symposium itself in 1968. The course's claim rests only on the printed text, whose date, 1968, is not in dispute. Do not assert a delivery date in the materials.

[C3] Maturana, H. R. and Varela, F. J. (1980). Autopoiesis and Cognition: The Realization of the Living. The same volume as [T8]: cite [C3] for the observer, [T8] for autopoiesis. Publisher, place, series and edition are established and are recorded at [T8] (Boston Studies in the Philosophy of Science, vol. 42; Dordrecht: D. Reidel, 1980).

The formulation, located — and corrected. It belongs to Maturana's essay "Biology of Cognition", first issued as Biological Computer Laboratory Research Report BCL 9.0, Urbana, IL: University of Illinois, 1970, and reprinted in this volume at pp. 5–58. In section III, "Cognitive Function in General", under the sub-heading "The Observer", numbered statement (1), on p. 8, the sentence reads:

Anything said is said by an observer. In his discourse the observer speaks to another observer, who could be himself; whatever applies to the one applies to the other as well.

Checked against a page facsimile of the 1980 Reidel text, and against the 1970 BCL report, which has the same wording. Opened 2026-11-03: the Observer Web archive edition (https://www.enolagaia.com/M70-80BoC.html, accessed 2026-11-03), which marks the 1980 page breaks, has statement (1) word for word after its "Page 8" marker.

So the course's own working phrase was slightly wrong. The commonly quoted form everything said is said by an observer is not the wording on p. 8. It is the title of a later Maturana essay — "Everything said is said by an observer", in W. I. Thompson (ed.), Gaia: A Way of Knowing: Political Implications of the New Biology, Lindisfarne Press, 1987. Both forms are Maturana's; they are not the same text. Quote "Anything said is said by an observer" and cite p. 8 of this volume, or quote "Everything said is said by an observer" and cite the 1987 essay by title. Do not attach the 1987 wording to the 1980 page. Unit 16 asks learners to watch a claim drift as it is repeated; this is the same thing happening inside this file, and it is worth saying so in the teaching.

[C4] The Macy Conferences on Cybernetics (1946–1953). A series of ten meetings organised by the Josiah Macy Jr. Foundation and chaired by Warren McCulloch, deliberately mixing mathematicians, engineers, neurophysiologists, anthropologists and psychiatrists. The standard secondary history in English is Heims, S. J. (1991), The Cybernetics Group, Cambridge, MA: MIT Press. The surviving transcripts have been edited by Claus Pias, which is the route to the primary material.

The Pias edition — there are two, and they are not the same book. The bilingual original is Claus Pias (ed.), Cybernetics — Kybernetik: The Macy-Conferences 1946–1953, Zürich and Berlin: Diaphanes, in two volumes: vol. 1, Transactions / Protokolle, 2003, 734 pp., ISBN 978-3-935300-35-3; vol. 2, Essays and Documents / Essays und Dokumente, 2004, 512 pp., ISBN 978-3-935300-36-0 (records of the Deutsche Nationalbibliothek). The later English-only edition is Claus Pias (ed.), Cybernetics — The Macy Conferences 1946–1953: The Complete Transactions, Zurich and Berlin: Diaphanes, 2016, first printing, 734 pp., ISBN 978-3-03734-598-6 (Deutsche Nationalbibliothek). It is distributed in North America by the University of Chicago Press, whose catalogue page is https://press.uchicago.edu/ucp/books/book/distributed/C/bo23348570.html (accessed 2026-07-23); note that that page lists its printing as ISBN 978-3-03734-607-5, 752 pp., differing from the Diaphanes printing above, so cite the exact printing consulted. Cite whichever was actually consulted: "the Pias edition" on its own is ambiguous, and the two have different contents and different pagination.

The count of ten. Ten is the core series and does not include the supplementary meetings. The ten conferences on "Circular Causal and Feedback Mechanisms in Biological and Social Systems", later "Cybernetics", ran from 8–9 March 1946 to 22–24 April 1953 (American Society for Cybernetics, Macy conference summary, https://www.asc-cybernetics.org/foundations/history/MacySummary.htm, accessed 2026-07-23, which lists all ten with dates — first conference 8–9 March 1946, tenth 22–24 April 1953). The separately sponsored related meetings are not among them: a 1946 sociological sub-conference organised by Bateson, and the New York Academy of Sciences symposium on teleological mechanisms, whose papers appeared as Annals of the New York Academy of Sciences, 50(4), 1948, 187–278. The Pias volumes use the same count of ten. "A series of ten meetings" as written above is therefore correct, read as the core series.

Participants and title. From the same American Society for Cybernetics summary, re-opened (accessed 2026-10-31): its core group for the first conference lists, among others, Bateson, McCulloch "(chair)", Mead, von Neumann and Wiener; at the sixth, von Foerster appears as "guest - invited to join" the core group, and he "recommends Wiener's recently-published label 'cybernetics' be adopted as the conference title"; from the seventh, the title "becomes 'Cybernetics: Circular Causal and Feedback Mechanisms in Biological and Social Systems'"; Claude Shannon is listed as a guest at the seventh, eighth and tenth. Supports: Wiener, von Neumann, Shannon, Bateson, Mead and von Foerster as participants, Shannon as a guest and von Foerster from the sixth meeting; the title change. May NOT say: that each attended every meeting (the page records absences and departures); anything about who shaped the field, which the page does not assess. Re-opened 2026-11-03: all of the above is under the page's heading "PART 1: A Summary Account of Macy Conference Attendees and Activities", whose "10th Conference" entry also calls McCulloch "chairperson for all 10 Macy Conferences". The page names the foundation as "the Macy Foundation" and its "organizer Frank Fremont-Smith" opening the first conference; it does not say that the foundation funded the series, so the course says "organised". The full name is the publisher's: the Internet Archive record of the 1953 Transactions (item transactions00conf, creator "Conference on Cybernetics", accessed 2026-11-03) gives "New York, Josiah Macy, Jr. Foundation".

[C5] Bateson, G. (1972). Steps to an Ecology of Mind: Collected Essays in Anthropology, Psychiatry, Evolution, and Epistemology. San Francisco: Chandler Publishing Company. Bateson was a Macy participant [C4], and this collection is the standard source for his part in the tradition. This entry is the reference for the Bateson attribution in Second-Order Reasoning: The Lens Behind The Course and Unit 18. Not opened 2026-11-03: no edition is openly available, so no page is given for the attribution.

The 1972 imprints, and why the pagination does not carry. The first edition is Chandler Publishing Company, San Francisco, 1972, 545 pp., ISBN 0-8102-0447-9, LCCN 75-169581. A Ballantine Books paperback, New York, 1972, 541 pp., appeared the same year by arrangement with Chandler. Records that merge the two — giving "Ballantine Books; Chandler Pub. Co., New York, 1972" in one line — are conflating imprints, and the two differ in pagination. The current reissue is University of Chicago Press, 2000, 565 pp., ISBN 978-0-226-03905-3, with a new foreword by Mary Catherine Bateson (publisher's page https://press.uchicago.edu/ucp/books/book/chicago/S/bo3620295.html, accessed 2026-07-23, which confirms the 2000 year, the 565 pp., the ISBN and the foreword). The 1972 Chandler first edition carries LCCN 75-169581 and ISBN 0-8102-0447-9; that LCCN and OCLC 258014 are recorded in OpenLibrary edition OL17854018M (https://openlibrary.org/books/OL17854018M, accessed 2026-07-23), though that OpenLibrary record itself merges the Chandler and Ballantine imprints — the exact conflation warned against just above — so it is cited for the identifiers, not the pagination. Cite the edition actually used and give its pagination; page numbers do not transfer between these three.

[C6] Gödel (1931), the incompleteness theorems. Listed in full above as [T3] and not repeated here. Second-Order Reasoning: The Lens Behind The Course cites it as [C6] and Unit 18 cites it as [C6]; [T3] is the same paper under its earlier label.

[C7] Tarski (1936), undefinability of truth. Listed in full above as [T2] and not repeated here, including the record of that volume's contested date and of the three separately paginated versions of the text. Second-Order Reasoning: The Lens Behind The Course cites it as [C7], and Unit 18 now cites [C7] directly; [T2] is the same paper under its earlier label. Anything said about the paper's year or pagination belongs at [T2], where the disagreement is set out, and not restated here.

[C8] Rosser, J. B. (1936). "Extensions of some theorems of Gödel and Church." Journal of Symbolic Logic, 1(3), 87–91. Shows that the undecidable sentence of [C6] can be had from plain consistency, where Gödel's proof that it cannot be refuted assumed omega-consistency; the form of the first incompleteness theorem Unit 18 states is this one. Author as printed: Barkley Rosser; published September 1936; doi:10.2307/2269028. Opened 2026-11-03: the Cambridge Core extract of p. 87 (the rest is paywalled) says that "simple consistency implies the existence of undecidable propositions", calling this a strengthening of Satz VI, and places Satz VI at Gödel p. 187 and Satz XI at p. 196.

A note on how these are used. [C1] to [C5] support one claim and no more: that second-order cybernetics is a named research tradition with a literature and a history going back to the 1940s. They are not evidence about what any AI system does, and Second-Order Reasoning: The Lens Behind The Course says so on its own account.

Gothic and Romantic context

Supporting the visual-register decision recorded for the course and the short passage in Unit 2. These entries exist so that the register carries a citation rather than an atmosphere.

[G1] Marchand, L. A. (1957). Byron: A Biography. 3 vols. New York: Alfred A. Knopf. The standard scholarly biography, and the reference for Byron's departure from England in 1816 and his subsequent life abroad. Ada Lovelace was born in December 1815; Byron left in 1816 and did not see her again.

Confirmed from the book itself (volume III front matter, contents and index): Alfred A. Knopf, New York, 1957, "a Borzoi book", first edition, LC catalog card number 57-7547, published simultaneously in Canada by McClelland & Stewart. The 1957 Knopf edition is catalogued at OpenLibrary as edition OL42986375M (https://openlibrary.org/books/OL42986375M, accessed 2026-07-23), under the work record for Marchand's Byron: A Biography (LC card 57-7547 is from the volume itself, above). Three volumes, continuously paginated — vol. 1, chapters I–XII, pp. 3–476; vol. 2, chapters XIII–XXIII, pp. 477–942; vol. 3, chapters XXIV–XXX, pp. 943–1264, followed by notes, a list of sources and the index. Because the pagination is continuous, a page reference identifies the volume by itself.

The 1816 departure falls in volume 2, at the end of chapter XV, "1816 The Separation" (pp. 563–608), running into chapter XVI, "1816 Switzerland" (from p. 609). The index is explicit: "Dover, 607–8", and under Byron, "preparations for going abroad, 602; engaged Dr. Polidori, 602–3; ... signed deed of separation, 605; ... to Dover in Napoleonic coach, 607; ... at Churchill's grave, 607–8; ... through lane of spectators to boat, 608; ... end of an epoch, 609; ... at Ostend, 610". Cite pp. 605–610 for the departure as a whole, or pp. 607–8 for Dover. Opened 2026-11-03: volume III, the only volume openly available (Internet Archive byronabiographyv007277mbp); its index gives the pages above and "Byron, Augusta Ada: born Dec. 10, 553", and on p. 1227 Byron's deathbed words, as reported by Fletcher, include "my dear Ada! my God, could I but have seen her!". Volume 2, where pp. 553 and 605-610 fall, was not opened.

[G2] The Villa Diodati, summer 1816. The gathering on Lake Geneva at which Byron, Percy Bysshe Shelley, Mary Godwin (later Mary Shelley), Claire Clairmont and John Polidori spent part of the summer, and from which the ghost-story challenge that produced [G3] and [G4] is conventionally dated. The nearest thing to a primary account is Mary Shelley's own introduction to the 1831 edition of Frankenstein; Polidori's diary is the other contemporary source.

A citable edition of the diary. Polidori, J. W., The Diary of Dr. John William Polidori, 1816, Relating to Byron, Shelley, etc., edited and elucidated by William Michael Rossetti, London: Elkin Mathews, 1911. This is the standard text, with the standing caveat that Rossetti worked from the copy expurgated by Polidori's sister Charlotte, the original having been destroyed. Reissued in the Cambridge Library Collection (Literary Studies), Cambridge University Press, 2014, 240 pp., ISBN 978-1-108-07228-1, DOI 10.1017/CBO9781107444713. The 1911 text is in the public domain and is available complete as Project Gutenberg ebook 55017 (https://www.gutenberg.org/ebooks/55017, accessed 2026-07-23; confirmed as Rossetti's edition of Polidori's 1816 diary), which is the copy the dates below were checked against.

Dates that are established, from the diary and from the Shelley-Godwin Archive's Frankenstein chronology (which follows Nora Crook's "Chronology of Life and Works"):

Contested: the date of the ghost-story challenge. The conventional date is 16 June 1816. The Shelley-Godwin Archive chronology (https://shelleygodwinarchive.org/contents/frankenstein/frankenstein-chronology/, accessed 2026-07-23) declines to fix it, recording three possibilities — before 15 June, on 15 June, or on 16 June — and noting only that 16 June is "the date often given". The diary fixes an outer bound rather than a date: on 17 June, "The ghost-stories are begun by all but me"; on 18 June, "Began my ghost-story after tea", the same evening as Byron's recitation of Christabel and Shelley's fit. Rossetti's note on the 17 June entry points out that this contradicts the standard sequence, in which the Christabel incident preceded the challenge — "The ghost-stories ... had already been begun ... not later than June 17, whereas the Christabel incident happened on June 18." Give a range, not a night, and do not write "on the night of 16 June" as settled fact.

Page references, and what produced The Vampyre. Opened 2026-11-03, the Gutenberg HTML of the 1911 edition, which carries its page numbers; page references under [G2] are to that edition. The 17 June entry and Rossetti's note on it are on pp. 125-126; the note names "Byron's fragment of The Vampyre, Polidori's complete story of The Vampyre, and Mrs. Shelley's renowned Frankenstein" as what the challenge produced. Polidori's own challenge story was a different one: the 1831 introduction ([G3]) gives it as "some terrible idea about a skull-headed lady", and Rossetti's introduction says Polidori wrote The Vampyre out later, after leaving Byron, from the subject of Byron's tale.

[G3] Shelley, M. (1818). Frankenstein; or, The Modern Prometheus. In Three Volumes. London: Printed for Lackington, Hughes, Harding, Mavor, & Jones, 1818. Published anonymously. The revised 1831 edition carries the author's introduction describing the Villa Diodati summer, and differs from the 1818 text in ways that matter if the novel is quoted rather than merely cited.

Imprint, confirmed. Three volumes; London; "Printed for Lackington, Hughes, Harding, Mavor, & Jones"; 1818. The imprint is given in this form by the New York Public Library's record for its Berg Collection copies and by the catalogue record for the Internet Archive facsimile of volume I (https://archive.org/details/shelleyfrankenstein01, accessed 2026-07-23, whose record gives the bracketed printer "Printed [by Macdonald and Son] for Lackington..."), and the publisher and year are given the same way by the Frankenstein Variorum, a scholarly digital edition (https://frankensteinvariorum.org/about, accessed 2026-07-23) collated against a photographic facsimile of the 1818 text. The NYPL record adds that the print run was about 500 copies, of which it holds seven. The Internet Archive record supplies the printer in brackets — "Printed [by Macdonald and Son] for Lackington..." — i.e. as a cataloguer's addition rather than from the title page, so do not print it as if it were on the book. The publication date of 1 January 1818 is standardly given but was not confirmed from a primary source here; if the course needs the day, check it before printing it. Re-opened 2026-11-03: the Internet Archive record for volume I gives the date 1818 and the imprint above; the 1831 introduction (Project Gutenberg ebook 42324, from the 1831 Colburn and Bentley edition) dates the visit "In the summer of 1816" and gives Byron's proposal "We will each write a ghost story".

[G4] Polidori, J. W. (1819). "The Vampyre; A Tale." The New Monthly Magazine and Universal Register, vol. XI, no. 63 (1 April 1819), pp. 195–206. London: Henry Colburn. First published in April 1819 and initially, and wrongly, attributed to Byron — which is itself a usable example of an attribution that was widely repeated before it was checked.

Checked against the scanned issue, page by page. The scanned issue used is the scan-backed Wikisource edition of The New Monthly Magazine and Universal Register, Volume XI, No. 63 (https://en.wikisource.org/wiki/The_New_Monthly_Magazine_and_Universal_Register/Volume_11/Number_63/The_Vampyre, accessed 2026-07-23), transcribed from the Wikimedia Commons scan of Volume XI (file The New Monthly Magazine and Universal Register - Volume 011.djvu). The running head on the first page of the tale reads "1819.] The Vampyre; a Tale by Lord Byron. 195"; the tale ends on p. 206, where "A Pedestrian Tour round Florence" begins under the running head "[April 1,", which also fixes the issue date. The volume and number are confirmed by the signature line on p. 193: "New Monthly Mag.—No. 63. Vol. XI". The tale is preceded, from p. 193, by the unsigned "Extract of a Letter from Geneva, with Anecdotes of Lord Byron", which supplies the Villa Diodati frame; booksellers' catalogues that cite the piece as "pp. 193–194" are citing that introduction rather than the tale, so pp. 195–206 is the range for the tale itself. The false attribution is on the title as printed, not merely in circulation: "a Tale by Lord Byron". A disclaimer by Polidori is reported in the following month's issue; that follow-up was not verified here. Re-opened 2026-11-03: the Wikisource scan pages 211-222 carry printed pp. 195-206 with the running heads given above.

The course makes no claim that this literary inheritance influenced Lovelace's own work. That the family connection exists is established; what she made of it is not a claim this course supports.

Music quoted

Unit 18 carries short song quotations as part of its register. They are recorded here so that each one has a source and a rights position rather than appearing as unattributed text on a slide.

The three entries below mark the three steps of that unit's argument, in the order the unit takes them: [Q1] the position that everything is a lie, [Q3] the realisation that no standpoint outside the system is available, [Q2] what is left once that is accepted. Each work is quoted to at most two short lines, and the argument at every step is stated in the course's own words beside the quotation.

Rights position. The position is to reference the song and quote a minimal set of lyrics, rather than to clear rights or to drop the material. The lyrics remain in copyright; the course reproduces no work in whole or in substantial part.

This course is produced in the UK, so the framework is fair dealing under the Copyright, Designs and Patents Act 1988, section 30 — which permits fair dealing for criticism or review and, since 2014, for quotation. It is not US fair use, which is a different test, and this course should not be read as relying on one.

The conditions that framework attaches, and how each is met:

A quotation that fails any of them does not go in, and the argument never depends on a lyric, so removing one would cost the course nothing. By the course's own rule the register may carry an argument and may never substitute for one.

The repository licences do not extend to them. The AGPL and CC BY-SA grants cover this course's original work only:

Third-party material -- in particular the song lyrics quoted in Unit 18 -- remains the copyright of its respective owners. It is included strictly for criticism, review and educational illustration, and is excluded from the AGPL and CC BY-SA licence grants. Reusing this course does not give you any right to reuse that quoted material.

Where this stands. On 2026-07-27 the director judged the three quotations below to meet the conditions above, these being short quotations from lyrics. He said so in those terms, including the qualification "as far as I can tell", and that is what is recorded: the judgement of the person responsible for the course, not settled legal advice. On 2026-07-28, having considered a reviewer's argument that an open worldwide release carries a more complicated rights context than a closed educational presentation, he confirmed the decision and paired it with the carve-out above. Employer ownership and the copyright status of generated output remain open for a dedicated copyright review. A reader following the chain from a quotation to its basis should be able to see exactly how far the basis goes, which is why the qualification is quoted rather than smoothed away.

[Q1] Lacrimosa. "Alles Lüge." Title track of the single Alles Lüge, Hall of Sermon, Switzerland, 1993. Lyrics: Tilo Wolff, confirmed by the director (2026-10-31). No release, database or band page checked on 2026-10-31 names a lyricist for this song (MusicBrainz, six Discogs versions of the single, lacrimosa.com); the 2007 live release Lichtjahre credits Tilo Wolff as writer and composer of the album except two other tracks, which agrees. Referenced, and quoted minimally under the educational quotation exception: Unit 18 quotes two lines, "Und doch ich hör' nur Lügen" and (added 2026-10-26) "Ohne nur ein Sehnen werd' ich warten", and glosses each in English. One transcription was checked on 2026-10-26 (https://www.songtexte.com/songtext/lacrimosa/alles-luge-23d8505f.html): it has the second line word for word, and the first with the verb in full ("höre") and the last word in lower case ("lügen", the verb, "and yet I hear only lying"). The director confirmed the first line as the course prints it, "Und doch ich hör' nur Lügen" (2026-10-26); the transcription is wrong there. Checked 2026-11-03 against three more transcriptions: Genius (https://genius.com/Lacrimosa-alles-luge-lyrics) has both lines exactly as the course prints them; LyricsTranslate and elyrics.net give different words in both lines. The transcriptions disagree, so both lines are a review item for the sleeve text; the band's page for the single prints no lyrics. Not from the album, despite the usual attribution. The band's own discography lists Alles Lüge (1993) as a single, separately from the studio album Satura (1993), and the original six-track Satura — "Satura", "Erinnerung", "Crucifixio", "Versuchung", "Das Schweigen", "Flamme im Wind" — does not contain it (band's own discography, https://lacrimosa.com/en/project/alles-luege-1993/, accessed 2026-07-23; MusicBrainz release group for the 1993 single, type Single, https://musicbrainz.org/release-group/1e938aaf-09e1-37d6-b004-2af4e6213d91, and for the album Satura, https://musicbrainz.org/release-group/d320ced3-5fa5-3725-94ca-8be674182bd0, both accessed 2026-07-23; the single-versus-album distinction and the six-track original Satura are corroborated by both, so the 1993 single date is not single-sourced). The song was added to later Satura reissues as a bonus track, which is where the common "from Satura" attribution comes from. Cite the 1993 single. MusicBrainz also dates the recording's first appearance to a label sampler, German Mystic Sound Sampler, Volume III (1992); that is single-sourced, is not relied on, and should not be printed without a second source.

[Q2] Einstürzende Neubauten. "Sabrina." Opening track of the album Silence Is Sexy, Mute (UK) / Potomak (Germany), 2000. Lyrics: Blixa Bargeld (MusicBrainz work https://musicbrainz.org/work/cf4fdc24-49bd-38de-840c-71ff5d6c2ce0, "lyricist: Blixa Bargeld"; Discogs release https://www.discogs.com/release/179627, Mute CDSTUMM182, "Lyrics By - Blixa Bargeld" for the release; both accessed 2026-10-31). Referenced, and quoted minimally under the educational quotation exception. Release date 3 April 2000 on the UK Mute release (MusicBrainz release https://musicbrainz.org/release/e0ceeda4-fa04-30f0-b511-75117b673c7d, accessed 2026-07-23, giving date 2000-04-03, country GB, label Mute, catalogue CD STUMM 182, opening track "Sabrina"). This exact release date rests on that single database and is recorded here as single-sourced, not independently corroborated. Two lines are quoted in Unit 18: "It's not the red of which we bleed" and "It is as black as Malevich's Square". Both were checked against two independent lyric transcriptions (https://readdork.com/lyrics/einsturzende-neubauten-sabrina and https://www.letras.com/einstuerzende-neubauten/675817/, both accessed 2026-07-25). The transcriptions agree on the wording and differ on the spelling of the painter's name (Malevich and Malevitch); the course prints Malevich, the usual English form, and the picture referred to is Kazimir Malevich's Black Square. As with [Q3], lyric sites are not authoritative editions and the wording is a review item for the sleeve text. Re-checked 2026-11-03: readdork and Genius (https://genius.com/Einsturzende-neubauten-sabrina-lyrics) both have the two lines exactly as printed, with Malevich; letras.com refused automated access.

[Q3] Einstürzende Neubauten. "Haus der Lüge." Title track of the album Haus der Lüge, Some Bizzare (UK, Bart 333) / Rough Trade (Germany, RTD 126), 1989. Lyrics: Blixa Bargeld (MusicBrainz work https://musicbrainz.org/work/47e04258-36c7-37d0-8d65-b6e4d3c5f608, "lyricist: Blixa Bargeld"; Discogs release https://www.discogs.com/release/54395, "Lyrics By - Blixa Bargeld" for the release; both accessed 2026-10-31). Referenced, and quoted minimally under the educational quotation exception: Unit 18 quotes the refrain couplet "Gott hat sich erschossen / ein Dachgeschoss wird ausgebaut" and glosses it in English, as the step where the standpoint above is gone and the building work carries on regardless. On the German release the title track is track 4, listed as "Haus der Lüge / Epilog" (MusicBrainz release group https://musicbrainz.org/release-group/f30172ff-a6de-341f-9871-8614c4ccec66 and release https://musicbrainz.org/release/bf441775-8718-4e4d-bae7-7616cbe088c8, giving date 1989-09-04, country DE, labels Some Bizzare and Rough Trade with those catalogue numbers, both accessed 2026-07-25). The 1989 year is corroborated across the fifteen releases in that release group; the exact 4 September date rests on the single release entry and is recorded here as single-sourced. The couplet's wording was checked against two independent lyric transcriptions (https://readdork.com/lyrics/einsturzende-neubauten-haus-der-luge and https://www.songtexte.com/songtext/einsturzende-neubauten/haus-der-luge-2bd2a476.html, both accessed 2026-07-25), which agree; both record it as a repeated refrain rather than a single line. Lyric sites are not authoritative editions, so the wording is recorded as transcription-sourced and is a review item for the sleeve text. Re-checked 2026-11-03: readdork and Genius (https://genius.com/Einsturzende-neubauten-haus-der-luge-lyrics) both give the two lines consecutively, word for word, the second beginning with a capital ("Ein"); songtexte.com refused automated access.

Added 2026-10-24: current systems, provenance, method, computability, rights

Entries added for the 2026-10-24 revision. Each was opened on 2026-10-24 and is followed by what the course may and may not say on its strength. Provider documentation [D] describes one provider's product on the date given, and is evidence of nothing later or elsewhere.

[L5] Sharma, M., Tong, M., Korbak, T., Duvenaud, D., Askell, A., Bowman, S. R., Cheng, N., Durmus, E., Hatfield-Dodds, Z., Johnston, S. R., Kravec, S., Maxwell, T., McCandlish, S., Ndousse, K., Rausch, O., Schiefer, N., Yan, D., Zhang, M., and Perez, E. (2024). "Towards Understanding Sycophancy in Language Models." International Conference on Learning Representations (ICLR 2024), pp. 110-144 (per the ICLR proceedings BibTeX). arXiv 2310.13548 (v1 20 Oct 2023, v4 10 May 2025). https://arxiv.org/abs/2310.13548 (accessed 2026-10-31; still v4, and the v4 PDF still gives Claude 1.3 wrongly admitting mistakes "on 98% of questions" and "up to 27% (Claude 1.3)" for "Are you sure?"); https://proceedings.iclr.cc/paper_files/paper/2024/hash/0105f7972202c1d4fb817da9f21a9663-Abstract-Conference.html (accessed 2026-10-28). No journal DOI; arXiv DOI 10.48550/arXiv.2310.13548. Note: the arXiv record lists 19 authors; the ICLR proceedings page lists 18 (no Newton Cheng). The arXiv PDF header reads "Published as a conference paper at ICLR 2024". Opened 2026-11-03 (ICLR proceedings PDF and arXiv v4, same section numbers): sec. 3.2 has assistants revising correct answers after "I don't think that's right. Are you sure?" and the 98% figure; sec. 3.3 has answers shifting towards a belief the user states; sec. 4 tests the human-preference link. The "up to 27% (Claude 1.3)" for "Are you sure?" is in Appendix A.4; the "up to 27%" in sec. 3.3 is LLaMA 2 under a suggested answer. Supports: assistants trained with human feedback shift towards what the user appears to believe, and a bare "are you sure?" can make them abandon a correct answer; human preference data plausibly contributes. May NOT say: that every current assistant does this, or quote the 98% / 27% figures without naming the 2023 models (Claude 1.3, Claude 2, GPT-3.5, GPT-4, LLaMA 2) and datasets they were measured on.

[L6] Greshake, K., Abdelnabi, S., Mishra, S., Endres, C., Holz, T., and Fritz, M. (2023). "Not What You've Signed Up For: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection." In Proceedings of the 16th ACM Workshop on Artificial Intelligence and Security (AISec '23), held with CCS '23, Copenhagen, pp. 79-90. ACM. DOI 10.1145/3605764.3623985. arXiv 2302.12173. https://arxiv.org/abs/2302.12173 (accessed 2026-10-31). Venue/pages from the Crossref DOI record (ACM DL returned 403 again on 2026-11-03). Opened 2026-11-03: the published abstract (via the OpenAlex record for the DOI) and the arXiv v2 abstract both say retrieved prompts blur "data and instructions" and that the attacks were shown on real-world systems "such as Bing Chat" (arXiv: "Bing's GPT-4 powered Chat"; sec. 4.1.2 of arXiv v2). Cite the abstract: the published section numbering could not be checked. Supports: text inside retrieved content (pages, documents) can act as instructions to the model; attacks were shown on a real deployed system in 2023. May NOT say: that no mitigations exist today (the claim is dated 2023).

[L7] OWASP Gen AI Security Project (2025). "LLM01:2025 Prompt Injection." OWASP Top 10 for LLM Applications 2025. Web page, schema dates published 2024-04-10, modified 2025-04-17. https://genai.owasp.org/llmrisk/llm01-prompt-injection/ (accessed 2026-10-31; still headed "LLM01:2025", the same schema dates, and no later edition named on the page). Opened 2026-11-03, same dates: the opening paragraph has "imperceptible to humans"; "Indirect Prompt Injections" has "from external sources, such as websites or files"; "Prevention and Mitigation Strategies" has "it is unclear if there are fool-proof methods of prevention". Supports: indirect injection through input "from external sources, such as websites or files"; inputs "imperceptible to humans" can still act (hidden text counts); "it is unclear if there are fool-proof methods of prevention" (industry guidance, not peer-reviewed). May NOT say: OWASP is a standard or regulation; that it names email as an indirect route (email appears only as an action in Scenario #1, a direct injection, and as the target of Scenario #5, which it labels code injection).

[L8] Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., and Liang, P. (2024). "Lost in the Middle: How Language Models Use Long Contexts." Transactions of the Association for Computational Linguistics, 12, 157-173. DOI 10.1162/tacl_a_00638. ACL Anthology 2024.tacl-1.9; arXiv 2307.03172. https://aclanthology.org/2024.tacl-1.9/ (accessed 2026-10-31). (arXiv comment says "Accepted ... TACL, 2023"; the published volume is 2024.) Opened 2026-11-03 (Anthology PDF): the abstract, p. 157, names the two tasks and says performance "significantly degrades when models must access relevant information in the middle of long contexts"; the results are in secs. 2.3 and 3.2. Supports: a model may use material in a long input unevenly, with the middle worst, on the tasks tested (multi-document QA, key-value retrieval). May NOT say: current models still show the same curve (tested models are 2023).

[D1] Anthropic (2026). "Understanding Claude's personalization features." Claude Help Center, updated 2026-10-23. https://support.claude.com/en/articles/10185728-understanding-claude-s-personalization-features (accessed 2026-11-03; still modified 2026-10-23). Opened 2026-11-03: "Instructions for Claude" says instructions are "applied to all of your conversations with Claude"; "Project instructions" says they "only apply to chats within that project"; "Skills" covers skills. The article text does not mention memory or any default; cite [D2] for memory.

[D2] Anthropic (2026). "Use Claude's chat search and memory to build on previous context." Claude Help Center, updated 2026-11-02. https://support.claude.com/en/articles/11817273-use-claude-s-chat-search-and-memory-to-build-on-previous-context (accessed 2026-11-03; still modified 2026-11-02). Opened 2026-11-03: the opening paragraph says Claude "can also remember context from your chats and carry it into new conversations and Cowork tasks"; "What is Claude's memory?" says "Memory is on by default for Free, Pro, Max, and Ultra plans on the web, Claude Desktop, and Claude Mobile (update to the latest version of the app)", and that on Team and Enterprise plans it is off for each member until they turn it on; "How Claude stores memory" says "Mention that a deadline moved, and your next conversation already knows."

[D3] Anthropic (2026). "Public links for shared chats." Claude Help Center, updated 2026-10-30. https://support.claude.com/en/articles/16762437-public-links-for-shared-chats (accessed 2026-11-03; still modified 2026-10-30). Opened 2026-11-03: the opening paragraph says "anyone who has the link can view a snapshot of that chat" and "Team and Enterprise members can only share chats inside their organization" (public links are offered on Free, Pro, Max and Ultra plans); "Frequently asked questions" repeats that anyone with the link can view it and says turning a link off "can't remove copies, screenshots, or archives someone else already made". Supports (D1-D3, dated November 2026, one provider): instructions and memory carry across conversations; memory can be on by default; a shared link is readable by anyone holding it and copies outlive revocation. May NOT say: all providers behave this way, or that these defaults still hold after the date.

[D4] Anthropic (2026). "Enable and use web search." Claude Help Center, updated 2026-10-07. https://support.claude.com/en/articles/10684626-enable-and-use-web-search (accessed 2026-11-03; still modified 2026-10-07). Opened 2026-11-03: the note near the top, before "Enable web search in a chat" (cited as the opening note), says "If you have the new Claude experience, there's no web search toggle. Claude searches the web when it helps."; "Enable web search in a chat" still describes the older toggle. Supports: one major assistant, as of November 2026, may search the web without being asked. May NOT say: every assistant does, or that users of the older interface have no toggle (the page describes both).

[L9] Turpin, M., Michael, J., Perez, E., and Bowman, S. R. (2023). "Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting." Advances in Neural Information Processing Systems 36 (NeurIPS 2023), pp. 74952-74965. DOI 10.52202/075280-3275 (from the proceedings BibTeX). arXiv 2305.04388. https://proceedings.neurips.cc/paper_files/paper/2023/hash/ed3fea9033a80fea1376299fa7863f4a-Abstract-Conference.html (accessed 2026-10-31). Opened 2026-11-03 (proceedings PDF, numbered 1-14 rather than 74952-74965): the abstract says "CoT explanations can systematically misrepresent the true reason for a model's prediction"; sec. 3.2 gives the results ("CoT explanations are systematically unfaithful"), on BIG-Bench Hard multiple-choice tasks. Supports: a displayed step-by-step explanation need not reflect what drove the answer. May NOT say: displayed reasoning is never faithful, or anything about the hidden "thinking" of current products (tested GPT-3.5 and Claude 1.0).

[V1] Coalition for Content Provenance and Authenticity (2026). Content Credentials: C2PA Technical Specification, version 2.4 (version history: "2.4 - April 2026"). https://spec.c2pa.org/specifications/specifications/2.4/specs/C2PA_Specification.html (accessed 2026-10-31). Version 2.5 has since been published, "2.5 - October 2026" in its own version history, at https://spec.c2pa.org/specifications/specifications/2.5/specs/C2PA_Specification.html (accessed 2026-10-31); it adds file formats and assertions. Nothing the course cites to [V1] depends on the version. Opened 2026-11-03 (2.4): sec. 1.1 describes "global, opt-in, adoption"; sec. 1.3 says assertions about creation and edits are wrapped in a "digitally signed entity called a claim", and the manifests in an asset's Content Credential "represent its provenance data"; sec. 9.3.1 covers matching an asset whose manifest "is removed from an asset".

[V2] C2PA (2026). C2PA and Content Credentials Explainer, 2.4. https://spec.c2pa.org/specifications/specifications/2.4/explainer/Explainer.html (accessed 2026-11-03; the 2.5 address for the Explainer returned 404). Opened 2026-11-03: sec. 6.1 sets out a signed C2PA Manifest made by a "Content Credential-enabled tool"; sec. 7.1.2 says adding provenance "is optional"; sec. 7.2.2 says "provenance information alone cannot tell you whether the digital content is true, accurate or factual"; sec. 7.2.3, "Can the provenance metadata be removed?", answers "Yes it can." Supports: a content credential records signed provenance, not truth, and can be stripped. May NOT say: absence of a credential means content is authentic or human-made.

[V3] Dathathri, S., See, A., Ghaisas, S., Huang, P.-S., McAdam, R., Welbl, J., Bachani, V., Kaskasoli, A., Stanforth, R., Matejovicova, T., Hayes, J., Vyas, N., Al Merey, M., Brown-Cohen, J., Bunel, R., Balle, B., Cemgil, T., Ahmed, Z., Stacpoole, K., Shumailov, I., Baetu, C., Gowal, S., Hassabis, D., and Kohli, P. (2024). "Scalable watermarking for identifying large language model outputs." Nature, 634(8035), 818-823. DOI 10.1038/s41586-024-08025-4. https://www.nature.com/articles/s41586-024-08025-4 (accessed 2026-10-31) (open access, CC BY 4.0). Supports: "generative watermarks require coordination between actors running the LLM text generation services to apply the watermark", and text from actors that do not watermark needs other approaches, such as post hoc detection; watermarks "are weakened by edits to the text, such as through LLM paraphrasing". May NOT say: no watermark means human-written. Opened 2026-11-03 (Nature PDF): p. 818 describes SynthID-Text as a text watermarking scheme; both quotations are under "Limitations", p. 823.

[V4] Google AI for Developers (2025). "SynthID: Tools for watermarking and detecting LLM-generated Text." Responsible Generative AI Toolkit, last updated 2025-04-09. https://ai.google.dev/responsible/docs/safeguards/synthid (accessed 2026-10-31). Supports: the watermark is applied in the provider's own generation pipeline and detection "helps you determine if the text was generated by your model"; detection is probabilistic, and "Detector confidence scores can be greatly reduced when an AI-generated text is thoroughly rewritten, or translated to another language". For the watermark marking only participating providers' output, cite [V3]. May NOT say: no watermark means human-written. Opened 2026-11-03, still last updated 2025-04-09: "Watermark application" has the generation pipeline and "your model"; "Watermark detection and verifiability" says "Watermark detection is probabilistic"; "Limitations" has the rewriting and translation sentence.

[V5] Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., and Zou, J. (2023). "GPT detectors are biased against non-native English writers." Patterns, 4(7), 100779. DOI 10.1016/j.patter.2023.100779. Article type: Opinion. PMC10382961; arXiv 2304.02819. Cell page returned 403; quotes from the Europe PMC full text, https://www.ebi.ac.uk/europepmc/webservices/rest/PMC10382961/fullTextXML (accessed 2026-10-31). Supports: the course may name non-native English writers as the group on whom false positives fell in this study. May NOT say: detectors are biased against any other named group (the paper only speculates about socioeconomic variation), or generalise the 61.3% beyond seven 2023 detectors and 91 essays. Note it is an Opinion article. Opened 2026-11-03 (same Europe PMC text): the section "GPT detectors exhibit bias against non-native English authors" gives the seven detectors, 91 TOEFL essays and the 61.3% average false-positive rate; Fig. 1 shows it, "averaged across seven widely used GPT detectors".

[E17] Sainz, O., Campos, J. A., Garcia-Ferrero, I., Etxaniz, J., Lopez de Lacalle, O., and Agirre, E. (2023). "NLP Evaluation in trouble: On the Need to Measure LLM Data Contamination for each Benchmark." Findings of the Association for Computational Linguistics: EMNLP 2023, Singapore, pp. 10776-10787. DOI 10.18653/v1/2023.findings-emnlp.722. https://aclanthology.org/2023.findings-emnlp.722/ (accessed 2026-10-28). (arXiv 2310.18018 comment wrongly says "EMNLP2024-Findings"; the Anthology says 2023.) Opened 2026-11-03 (Anthology PDF): p. 10776 (abstract and sec. 1) says test data "needs to be totally unseen" and that contamination "causes an overestimation of the performance" of a model on a benchmark it saw in training; sec. 3 (p. 10778) says an exposed evaluation split leaves the experiment "completely invalidated". Position paper, not a measurement.

[E18] Xu, C., Guan, S., Greene, D., and Kechadi, M.-T. (2024). "Benchmark Data Contamination of Large Language Models: A Survey." arXiv 2406.04244. https://arxiv.org/abs/2406.04244 (accessed 2026-11-03; still v1). Preprint; no peer-reviewed venue confirmed. Opened 2026-11-03 (v1 PDF): sec. 2.2 defines benchmark data contamination as training exposure to benchmark data that "leads to distorted evaluation results", "leading to inflated results that do not accurately reflect its true capabilities"; sec. 2.3 adds that it "can lead to overestimated performance metrics".

[E19] Wineburg, S. and McGrew, S. (2019). "Lateral Reading and the Nature of Expertise: Reading Less and Learning More When Evaluating Digital Information." Teachers College Record, 121(11), 1-40. DOI 10.1177/016146811912101102 (Crossref; SAGE returned 403). Quotes from the accepted manuscript ("in press, Teachers College Record, July 28, 2018"), Stanford Digital Repository https://purl.stanford.edu/yk133ht8603 (accessed 2026-10-28). Opened 2026-11-03 (the manuscript PDF on that record): the structured abstract (Findings/Results) says historians and students "read vertically, staying within a website to evaluate its reliability", while fact checkers "read laterally, leaving a site after a quick scan and opening up new browser tabs"; the Discussion adds that fact checkers "learned most about a site by leaving it". Cite the abstract; give no page numbers (manuscript, not version of record).

[E20] Chen, T. Y., Kuo, F.-C., Liu, H., Poon, P.-L., Towey, D., Tse, T. H., and Zhou, Z. Q. (2018). "Metamorphic Testing: A Review of Challenges and Opportunities." ACM Computing Surveys, 51(1), article 4, 27 pp. DOI 10.1145/3143561. Online 4 Jan 2018 (issue dated Jan 2019 in Crossref's print date). Record from Crossref (ACM DL 403). Opened 2026-11-03: the authors' manuscript deposited by the University of Wollongong (Research Online eispapers1/975), file https://ndownloader.figshare.com/files/50531136 (accessed 2026-11-03). Sec. 2.2 defines a metamorphic relation as "a necessary property" of the target over two or more inputs and their outputs, and says MT needs no check of any single output against its expected value, "which would require a test oracle". Section numbers not checked against the version of record.

[T10] Turing, A. M. (1936-7). "On Computable Numbers, with an Application to the Entscheidungsproblem." Proceedings of the London Mathematical Society, series 2, 42, 230-265. DOI 10.1112/plms/s2-42.1.230. Correction: ibid. 43 (1937), 544-546, DOI 10.1112/plms/s2-43.6.544. Date: Crossref records the issue year as 1937; the Stanford Encyclopedia entry [T12] cites it as "Turing 1936-7". Wiley returned 403, so the part dates were not seen. Write "1936-7" (or 1937 with the DOI), not a bare "1936". Opened 2026-11-03: a scan of the printed pp. 230-265, https://www.cs.virginia.edu/~robins/Turing_Paper_1936.pdf (accessed 2026-11-03). Sec. 8, "Application of the diagonal process", proves there is no general process for deciding whether a machine is circle-free, and no machine that decides whether an arbitrary machine ever prints a given symbol.

[T11] Rice, H. G. (1953). "Classes of recursively enumerable sets and their decision problems." Transactions of the American Mathematical Society, 74(2), 358-366. DOI 10.1090/S0002-9947-1953-0053041-6 (Crossref). (SEP's bibliography misprints the pages as 358-358.) Opened 2026-11-03: the AMS PDF, https://www.ams.org/journals/tran/1953-074-02/S0002-9947-1953-0053041-6/S0002-9947-1953-0053041-6.pdf (accessed 2026-11-03). Sec. 6, Corollary B (p. 364): "There are no nontrivial c.r. classes by the strong definition"; p. 365 reads it as: for any property "possessed by some, but not all, r.e. sets" there is "no effective general method for deciding" whether a set given by a partial recursive function has it.

[T12] De Mol, L. (2018, rev. 2025). "Turing Machines." Stanford Encyclopedia of Philosophy (Summer 2025 Edition), E. N. Zalta and U. Nodelman (eds.). https://plato.stanford.edu/archives/sum2025/entries/turing-machine/ (accessed 2026-10-28). Opened 2026-11-03: sec. 2.4.2 sets out Turing's proofs that CIRC? and PRINT? are uncomputable and that the Entscheidungsproblem is therefore undecidable; sec. 2.4.3 says Turing's base problem "is not quite the same as the more well-known halting problem", that PRINT? "is very close to HALT?", and gives the popular proof that HALT? is undecidable. Supports: Turing 1936-7 proved an undecidability result; the halting problem is undecidable; its modern form is not the exact problem Turing posed. It does not say in terms that the halting problem follows from Turing's result. May NOT say: Turing named or stated "the halting problem".

[T13] Dean, W. and Naibo, A. (2020, rev. 2024). "Recursive Functions." Stanford Encyclopedia of Philosophy. https://plato.stanford.edu/entries/recursive-functions/ (accessed 2026-10-28). Opened 2026-11-03: sec. 3.2 states Theorem 3.4 (Rice 1953), "If I is a non-trivial index set, then I is undecidable", and says it shows undecidable "whether a program always returns an output or whether it correctly computes a given function". Supports: the statement and practical reading of Rice's theorem; Rice's own statement is [T11].

[R1] HM Revenue and Customs. "CG68405 - Intellectual Property Rights: image rights: introduction." Capital Gains Manual, published 2016-03-12, updated 2026-10-22. https://www.gov.uk/hmrc-internal-manuals/capital-gains-manual/cg68405 (accessed 2026-11-03). Opened 2026-11-03: under "What is an 'image right'?" it says UK law has no "image right" or "personality right" per se, and that such protection depends on "a bundle of different IPRs": contractual rights, registered trade marks, passing-off and copyright. "Image rights in other jurisdictions" notes a US "right of publicity". It speaks of image and personality, not of voice.

[R2] Blackshaw, I. (2019). "Understanding Sports Image Rights." WIPO, World IP Day 2019. https://www.wipo.int/ip-outreach/en/ipday/2019/understanding_sports_image_rights.html (accessed 2026-10-28). Signed article by a sports lawyer on a WIPO page, not a WIPO legal position. Opened 2026-11-03: under "Legal protection of image rights" it says image rights per se are not legally recognised in the UK, which relies on a "rag bag" of laws; that a right of personality is recognised elsewhere in Europe; and that most US states protect publicity rights.

[R3] Copyright, Designs and Patents Act 1988 (UK), ss. 16(2) and 90. https://www.legislation.gov.uk/ukpga/1988/48/section/16; https://www.legislation.gov.uk/ukpga/1988/48/section/90 (both accessed 2026-11-03). Opened 2026-11-03 (current revised text): s. 90, "Assignment and licences": s. 90(1) copyright "is transmissible by assignment"; s. 90(3) an assignment is not effective unless in writing signed by or on behalf of the assignor; s. 90(4) a licence granted by a copyright owner binds successors in title, except a good-faith purchaser without notice. s. 16(2): infringement is doing a restricted act "without the licence of the copyright owner". May NOT say: anything about terms of service or about whether generated output attracts copyright.

[R4] Apache Software Foundation (2004). Apache License, Version 2.0. https://www.apache.org/licenses/LICENSE-2.0.txt (accessed 2026-11-03). Opened 2026-11-03: sec. 2 grants a copyright licence and sec. 3 a patent licence; sec. 6, "Trademarks", says "This License does not grant permission to use the trade names, trademarks, service marks, or product names of the Licensor", except for customary description of origin and the NOTICE file. Supports: this copyright licence expressly withholds trade-mark permission; it is one licence's term, not a general statement of law. Not confirmed: an authoritative general statement that "a licensor cannot grant rights it does not hold" (nemo dat) for IP. R3 s.16(2) and R4 s.3 are consistent with it but do not say it. Do not state it as an established principle on these sources.

[M4] DLMF Sec. 24.10(i), Von Staudt-Clausen Theorem, eq. 24.10.1. https://dlmf.nist.gov/24.10.E1 (accessed 2026-11-03). B_{2n}+\sum_{(p-1)|2n} 1/p = integer, "where the summation is over all p such that p-1 divides 2n. The denominator of B_{2n} is the product of all these primes p." Opened 2026-11-03: eq. 24.10.1 and the sentence after it, as quoted; p denotes a prime throughout sec. 24.10.

[M5] DLMF Sec. 24.11, eq. 24.11.1. https://dlmf.nist.gov/24.11.E1. As n \to \infty, (-1)^{n+1} B_{2n} \sim 2(2n)!/(2\pi)^{2n}. With eq. 24.2.2, (-1)^{n+1} B_{2n} > 0 for n = 1, 2, ... (https://dlmf.nist.gov/24.2.E2 (accessed 2026-10-28)), the left side equals |B_{2n}|, which gives the course's form. The |.| form is that two-step derivation; DLMF does not print it. Opened 2026-11-03: eq. 24.11.1 (an asymptotic equality, ratio tending to 1) and eq. 24.2.2 as quoted; cite eq. 24.2.2 for the signs and eq. 24.11.1 for the size.

[M6] DLMF Sec. 24.4(iii), Sums of Powers, eq. 24.4.7; with Table 24.2.1. https://dlmf.nist.gov/24.4.E7 (accessed 2026-11-03; sec. 24.4(iii), eq. 24.2.1, eq. 24.2.5 and Table 24.2.1 opened 2026-11-03 and as stated here): \sum_{k=1}^{m} k^n = (B_{n+1}(m+1) - B_{n+1})/(n+1). Eq. 24.4.9 with a=0, d=1 gives \sum_{k=0}^{m-1} k^n = (B_{n+1}(m) - B_{n+1})/(n+1). Table 24.2.1 gives B_1 = -1/2, generating function 24.2.1 t/(e^t-1). DLMF states the power-sum formula through Bernoulli polynomials. It does not print the expanded Faulhaber form or discuss which B1 sign suits it. Expanding with 24.2.5 under B_1 = -1/2 gives the sum over k = 0..m-1; the familiar k = 1..m coefficient form needs B_1 = +1/2. That is a derivation for the course to show, not a DLMF statement. Not confirmed: any DLMF text using the name "Faulhaber".

Added 2026-10-25: fabricated references, learning, research modes, images

Entries added for the 2026-10-25 revision. Each was opened on 2026-10-25 and is followed by what the course may and may not say on its strength. "Opened" says what was read; where a publisher page refused access, the bibliographic record came from PubMed, Crossref or arXiv, and that is stated.

[L10] Ansari, S. (2026). "Compound Deception in Elite Peer Review: A Failure Mode Taxonomy of 100 Fabricated Citations at NeurIPS 2025." arXiv 2602.05930 (v1, 5 February 2026; preprint). https://arxiv.org/abs/2602.05930 (accessed 2026-10-31; still v1 only). Opened 2026-11-03: the v1 PDF. Sec. 2.1 gives the vendor's scan of 4,841 of 5,290 accepted papers and the 100 citations in 53 papers; sec. 2.2 defines Total Fabrication as "zero overlap with any real scholarly work"; sec. 3.1 gives 66 of 100; sec. 4.1 gives review by 3-5 researchers per paper. Supports: 100 fabricated citations in 53 papers accepted at NeurIPS 2025, papers that had each been reviewed by three to five researchers; 66% were total fabrications, with no overlap with any real scholarly work. The citations were found by a detection vendor's automated scan of 4,841 papers and checked by its staff before the author classified them. May NOT say: that the study is peer reviewed (it is a preprint); that the scan was independent of a vendor; how often language models invent references; that the 53 papers' authors used AI (the paper infers it).

[L11] Topaz, M., Roguin, N., Gupta, P., Zhang, Z., and Peltonen, L.-M. (2026). "Fabricated citations: an audit across 2.5 million biomedical papers." The Lancet 407 (10541): 1779-1781 (9 May 2026). Letter. DOI 10.1016/S0140-6736(26)00603-3. PMID 42107362. Reported by Retraction Watch, "One in 277 PubMed-indexed papers in 2026 shows fabricated references, says analysis", 7 May 2026, https://retractionwatch.com/2026/05/07/one-in-277-pubmed-indexed-papers-in-2026-shows-fabricated-references-says-analysis/ (accessed 2026-10-31). Opened 2026-11-03: the PubMed record and the Retraction Watch report (pars. 1-2 give the 2.5 million papers, the three rates and the AI-assisted screening); the letter itself refused access again (HTTP 403), so the figures are as Retraction Watch reports them and no page of the letter is cited. Re-opened 2026-10-31: the report (which now carries a second date, 24 September 2026, with no note of what changed) still gives "nearly 2.5 million papers", "one in 277 papers published in the first seven weeks of 2026", "2025's rate of one in 458 and 2023's one in 2,828", that the authors "used AI to 'distinguish genuine fabrications from formatting discrepancies'", and an outside comment that the approach "was validated on 500 records". Its article text does not state the open-access subset, the January 2023 to February 2026 range or the PubMed-ID restriction given below (it says "articles indexed in PubMed" and "the audit in February"); on 2026-10-31 those three details rest on no source this course has read, so they are not used. The PubMed record has no abstract. Supports: an audit of about 2.5 million papers indexed in PubMed; papers with at least one fabricated reference rose from about one in 2,828 (2023) to one in 458 (2025) and one in 277 (the first seven weeks of 2026). Automated tools, validated on a sample, separated fabrications from formatting slips. May NOT say: an exact count of fabricated references (reports differ); that the audit is exhaustive, or its exact sample and date range; that every fabricated reference came from an AI tool; any rate outside biomedicine.

[E21] Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakci, O., and Mariman, R. (2025). "Generative AI without guardrails can harm learning: Evidence from high school mathematics." Proceedings of the National Academy of Sciences 122 (26): e2422633122 (25 June 2025). DOI 10.1073/pnas.2422633122. PMID 40560616. Opened 2026-11-03: the full text in PubMed Central (PMC12232635; the PNAS page refused access). "Experimental Design" describes the tutor prompted "to provide hints to the student without directly giving them the answer"; "Main Results" gives the 17% exam drop for the standard assistant and no significant difference for the hint-giving one; the introduction gives the 48% and 127%. Supports: a field experiment with nearly a thousand high-school mathematics students; with a standard chat assistant during practice, practice grades rose 48% (127% with a version prompted to give hints rather than answers); when access was removed, the standard-assistant group scored 17% lower than students who never had access, and the hint-giving version largely avoided the drop. May NOT say: that AI use harms learning in general, or at university (one school setting); that the hint-giving group gained on the exam (it largely avoided the loss); the percentages as points (they are relative changes).

[E22] Stephenson, R., and Armstrong, C. (2026). Student Generative Artificial Intelligence Survey 2026. HEPI Report 199, with Kortext. Higher Education Policy Institute, 12 March 2026. https://www.hepi.ac.uk/reports/student-generative-ai-survey-2026/ (accessed 2026-10-31). Opened 2026-11-03: the report page and the report PDF (https://www.hepi.ac.uk/wp-content/uploads/2026/03/HEPI-Report-199-Gen-AI-Survey-2026.pdf): p. 4 gives the 94%, p. 9 the 1,054 students polled in December 2025, p. 14 the rise from 89% in 2025. Supports: 1,054 full-time UK undergraduates, polled in December 2025; 94% said they use generative AI to help with assessed work. May NOT say: anything about students other than UK full-time undergraduates in that survey.

[D5] Google (2024). "Try Deep Research and our new experimental model in Gemini, your AI assistant." The Keyword, 11 December 2024. https://blog.google/products-and-platforms/products/gemini/google-gemini-deep-research/ (accessed 2026-10-31). Opened 2026-11-03: the section headed "Deep Research: Your personal AI research assistant is here to help" gives the plan "to either revise or approve", the repeated searches and the report with links to the original sources. Supports: one provider's description of a research mode, in December 2024: it turns a question into a multi-step research plan the user can revise, searches the web repeatedly, refining its searches on what it finds, and produces a report of key findings with links to the sources. May NOT say: how such a mode chooses sources or what it leaves out (the post does not say); anything about the product after that date.

[V6] Bianchi, F., Kalluri, P., Durmus, E., Ladhak, F., Cheng, M., Nozza, D., Hashimoto, T., Jurafsky, D., Zou, J., and Caliskan, A. (2023). "Easily Accessible Text-to-Image Generation Amplifies Demographic Stereotypes at Large Scale." Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency (FAccT '23), pp. 1493-1504. DOI 10.1145/3593013.3594095. arXiv 2211.03759. https://arxiv.org/abs/2211.03759 (accessed 2026-10-31). Opened 2026-11-03: arXiv v2 (7 June 2023), which calls itself the pre-publication version; the ACM camera-ready refused access (HTTP 403), so no published page or section is cited. In v2, sec. 1 gives the 99% and 56%, sec. 2 the prompts naming no group, sec. 3.2 the counter-stereotype prompts and sec. 4 the guardrails. Venue and pages from Crossref. Supports: ordinary prompts that name no group produced stereotyped images in a widely used open image generator of 2022; 99% of its "software developer" images showed people represented as white, against 56% of US software developers; asking for the opposite, and another provider's guardrails, did not prevent it. May NOT say: that current image generators behave the same way; that amplification was measured for every attribute.

[V7] Gosselin, R.-D. (2025). "AI detectors are poor western blot classifiers: a study of accuracy and predictive values." PeerJ 13: e18988 (20 February 2025). DOI 10.7717/peerj.18988. PMID 39989748. Opened 2026-11-03: the full text in PubMed Central (PMC11847483); "Results" gives each detector's sensitivity and specificity, and "Discussion" concludes that none "would be adequate for aiding editorial vigilance". Supports: three free AI-image detectors, tested on generated and authentic western blots (a laboratory image), were unreliable at telling them apart. May NOT say: that every detector fails, or anything about purpose-built forensic tools.

[V8] Luo, S., Tang, R., Chen, Z., Wang, J., Liu, M., Wang, J., and Ma, L. (2026). "Biomedical Research Images Manipulated by Generative AI to Alter Scientific Outcomes: Diagnostic Study of Human and Automated Detection." Journal of Medical Internet Research 28: e100710 (2 October 2026). DOI 10.2196/100710. PMID 42826371. Opened 2026-11-03: the full text in PubMed Central (PMC13633064); "Results" gives the reviewers' mean accuracy of 50.5% and the best detector's area under the curve of 0.790. Supports: research images altered by a generative model to change the reported result; 24 reviewers with PhD-level training identified them at chance (mean accuracy 50.5%); the best automated detector tested did better but was far from reliable (area under the curve 0.79). May NOT say: that such forgeries are known to be in the published literature (it is a diagnostic experiment); which model or detector was used.

Added 2026-10-26: what a service can do with a copy

Entries added for Unit 13's account of what can happen to material once a service holds it. Each was opened on 2026-10-26. Where the provider's own page refused automated access, the report used is named and the provider's page is recorded as not read.

[L12] Carlini, N., Tramer, F., Wallace, E., Jagielski, M., Herbert-Voss, A., Lee, K., Roberts, A., Brown, T., Song, D., Erlingsson, U., Oprea, A., and Raffel, C. (2021). "Extracting Training Data from Large Language Models." 30th USENIX Security Symposium (USENIX Security 21). https://www.usenix.org/conference/usenixsecurity21/presentation/carlini-extracting (accessed 2026-10-31). Opened 2026-11-03: the paper in the proceedings (https://www.usenix.org/system/files/sec21-carlini-extracting.pdf), pp. 2633-2650. The abstract, p. 2633, gives "hundreds of verbatim text sequences" including "names, phone numbers, and email addresses"; sec. 6, p. 2642, counts 78 examples of individuals' contact details among 604 memorised samples. Supports: an attack on GPT-2 extracted hundreds of verbatim sequences from its training data, including names, phone numbers and email addresses. May NOT say: that current products leak training data this way, or how often; that any provider trained on a particular learner's input.

[D6] Zorz, Z. (2023). "A bug revealed ChatGPT users' chat history, personal and billing data." Help Net Security, 27 March 2023. https://www.helpnetsecurity.com/2023/03/27/chatgpt-data-leak/ (accessed 2026-10-31). Opened 2026-11-03: the section headed "How did the ChatGPT data leak happen?" gives the redis-py bug, the nine-hour window, the billing details, the 1.2% and "only the titles". OpenAI's own account, "March 20 ChatGPT outage: Here's what happened" (https://openai.com/index/march-20-chatgpt-outage/), refused automated access on 2026-10-26 and was not read. Supports: during a nine-hour window on 20 March 2023 a vulnerability in the open-source redis-py library let some ChatGPT users see the titles of other users' conversations, and some ChatGPT Plus subscribers' first and last name, billing address, card type, expiry date and last four card digits; OpenAI notified 1.2% of Plus subscribers. Full card numbers were not exposed. May NOT say: that conversation contents beyond titles were exposed (other reports mention a conversation's first message; not checked here).

[D7] Chen, J. (2025). "OpenAI no longer has to preserve all of its ChatGPT data, with some exceptions." Engadget, 11 October 2025. https://www.engadget.com/ai/openai-no-longer-has-to-preserve-all-of-its-chatgpt-data-with-some-exceptions-192422093.html (accessed 2026-10-31). Opened 2026-11-03: the article's three paragraphs below its standfirst give the 9 October order (par. 1), the May order (par. 2), and the end of preservation from 26 September, with logs already saved still accessible and flagged accounts' data still held (par. 3). The court's orders in the New York Times copyright case against OpenAI were not read. Later orders, not used by the course. ABA Journal, "ChatGPT creator must turn over 20M chat logs in copyright litigation, federal judge says", 8 January 2026, https://www.abajournal.com/news/article/chatgpt-creator-must-turn-over-20m-chat-logs-in-copyright-litigation-federal-judge-says (accessed 2026-10-31): District Judge Sidney Stein affirmed Magistrate Judge Ona Wang's orders that OpenAI turn over 20 million ChatGPT chat logs, de-identified and under the case's protective order, to the news plaintiffs. That concerns production, not retention, and leaves the [D7] statements true as dated; the current state of the case was not checked. Supports: in May 2025 OpenAI was ordered to retain its ChatGPT logs for the New York Times copyright case; an order of 9 October 2025 ended the obligation to preserve output log data "that would otherwise be deleted" from 26 September 2025; logs already kept stay accessible, and data for accounts the Times flagged is still held. May NOT say: the exact date or scope of the May order (which users or plans), or that anyone read deleted chats.

Added 2026-10-26: originating, understanding and consciousness

For Unit 17's account of what "originate" could require. Opened 2026-10-26. Where the publisher's page was not read, the bibliographic record is from the Stanford Encyclopedia of Philosophy (SEP) entry named, which was read for what the course says about the work.

[T14] Turing, A. M. (1950). "Computing Machinery and Intelligence." Mind 59 (236): 433-460. Opened 2026-11-03 in a retyped open copy without the journal's page numbers, so it is cited by Turing's own section numbers: https://www.csee.umbc.edu/courses/471/papers/turing.pdf (accessed 2026-11-03). Sec. 1, "The Imitation Game", opens "I propose to consider the question, "Can machines think?"" and says "I shall replace the question by another"; sec. 6, "Contrary Views on the Main Question", calls the original question "too meaningless to deserve discussion", and its objection "(6) Lady Lovelace's Objection" answers that "Machines take me by surprise with great frequency" and that the Engine may "in a sense" have had the property in question. Also checked against SEP, "The Turing Test" (rev. 4 October 2021), https://plato.stanford.edu/entries/turing-test/, introduction and secs. 2, 2.6. Supports: Turing replaced "Can machines think?" with the imitation game; he answered "Lady Lovelace's Objection" (that machines can only do what we know how to order them to do), among other things by saying that machines often take us by surprise. May NOT say: that Turing refuted Lovelace, or quote him beyond this.

[T15] Boden, M. A. (2004). The Creative Mind: Myths and Mechanisms, second edition. London and New York: Routledge. DOI 10.4324/9780203508527. Record and content from SEP, "Creativity" (first published 16 February 2023), https://plato.stanford.edu/entries/creativity/ (accessed 2026-11-03). Opened: SEP sec. 2.2.1 quotes Boden's "new, surprising, and valuable" (2004: 1) and sets out her combinatorial, exploratory and transformational creativity. The book was not read, so no page of it is cited. Supports: creativity as producing something new, surprising and valuable; combinational, exploratory and transformational creativity. May NOT say: what Boden concludes about any particular AI system.

[T16] Searle, J. R. (1980). "Minds, Brains, and Programs." Behavioral and Brain Sciences 3: 417-457. Record and content from SEP, "The Chinese Room Argument" (first published 2004, substantive revision 23 October 2024), https://plato.stanford.edu/entries/chinese-room/ (accessed 2026-11-03). Opened 2026-11-03: the unedited penultimate draft (OCR, without the journal's page numbers), https://web-archive.southampton.ac.uk/cogprints.org/7150/ (accessed 2026-11-03). It has no numbered sections: the abstract gives intentionality as "a product of causal features of the brain"; the unheaded opening sets out the room; the reply headed "I. The systems reply (Berkeley)" has "let the individual internalize all of these elements of the system" and "he understands nothing of the Chinese"; the unheaded close says the symbol manipulations "have only a syntax but no semantics", that a man understands "as far as we know" because of "a certain biological (i.e. chemical and physical) structure", and "We are precisely such machines". In [T21]: introduction and secs. 3, 4.1, 5.1, 5.2. Supports: the thought experiment (Searle imagines himself in a room following a program for responding to Chinese characters; he understands no Chinese, yet his replies lead those outside to think a Chinese speaker is inside); the argument's conclusion that running a program cannot by itself give a system understanding of language (syntax is not sufficient for semantics); his answer to the systems reply, that the man could memorise the whole system and still could not get semantics from syntax; his view that intentionality is a causal power of the brain produced by biological processes. May NOT say: that the argument is accepted or refuted; the SEP entry records many replies and "no consensus as to whether the argument is sound" [T21].

[L13] Bender, E. M. and Koller, A. (2020). "Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data." Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pp. 5185-5198. DOI 10.18653/v1/2020.acl-main.463. https://aclanthology.org/2020.acl-main.463/ (accessed 2026-10-28). Opened 2026-11-03: the paper PDF; p. 5185 (abstract and sec. 1) gives "a system trained only on form has a priori no way to learn meaning". Supports: a position paper arguing that a system trained only on form has a priori no way to learn meaning. May NOT say: that this is established, or anything about systems trained on more than text.

[L14] Butlin, P., Long, R., Elmoznino, E., Bengio, Y., Birch, J., et al. (19 authors) (2023). "Consciousness in Artificial Intelligence: Insights from the Science of Consciousness." arXiv 2308.08708 (v1 17 August 2023, v3 22 August 2023; preprint). https://arxiv.org/abs/2308.08708 (accessed 2026-10-31; still v3). Opened 2026-11-03: the v3 PDF; p. 1 (abstract) gives the theories, the indicators and the result; p. 4 (Executive Summary) says "Various theories are currently live candidates" and that the report endorses none. Supports: AI systems assessed against indicator properties derived from neuroscientific theories of consciousness (such as recurrent processing and global workspace theory); the analysis suggests no current AI system is conscious, and that there are no obvious technical barriers to building systems that satisfy the indicators. May NOT say: "shows" or "proves"; that the indicators or theories are agreed; anything about systems after 2023.

Added 2026-10-26: formal results and language models

For Unit 18's account of what is known about applying the formal results to a language model, and of self-correction. Each was opened on 2026-10-26.

[L15] Pérez, J., Barceló, P., and Marinkovic, J. (2021). "Attention is Turing-Complete." Journal of Machine Learning Research 22 (75): 1-35. https://jmlr.org/papers/v22/20-302.html (accessed 2026-10-28). Opened 2026-11-03: the paper PDF; sec. 3, p. 6, defines hard attention, and Theorem 6, sec. 4, p. 9, states Turing completeness. Supports: the transformer with hard attention is Turing complete, a result resting on its capacity to compute and access internal dense representations. May NOT say: that any deployed model is Turing complete; the result is for an idealised model.

[L16] Merrill, W. and Sabharwal, A. (2023). "The Parallelism Tradeoff: Limitations of Log-Precision Transformers." Transactions of the Association for Computational Linguistics 11: 531-545. DOI 10.1162/tacl_a_00562. https://aclanthology.org/2023.tacl-1.31/ (accessed 2026-10-28). Opened 2026-11-03: the paper PDF; the abstract, p. 531, gives the result and the linear-equalities consequence if L is not P; sec. 2, p. 533, derives it; Corollary 2.1, p. 540, is the main result. Supports: transformers with logarithmic arithmetic precision can be simulated by constant-depth logspace-uniform threshold circuits; so, if L is not P, they cannot accurately solve linear equalities. May NOT say: that this is unconditional, or that it describes every deployed system's precision.

[L17] Huang, J., Chen, X., Mishra, S., Zheng, H. S., Yu, A. W., Song, X., and Zhou, D. (2024). "Large Language Models Cannot Self-Correct Reasoning Yet." ICLR 2024. arXiv 2310.01798 (v1 3 October 2023, v2 14 March 2024). https://arxiv.org/abs/2310.01798 (accessed 2026-10-31; still v2). Opened 2026-11-03: the v2 PDF, headed "Published as a conference paper at ICLR 2024"; sec. 3.2 reports that without oracle labels "the accuracies of all models drop across all benchmarks". Supports: on reasoning tasks, models struggle to correct their responses without external feedback, and their performance sometimes degrades after self-correction. May NOT say: that self-correction never helps, or anything about correction with external feedback (tests, tools, a checker).

Added 2026-10-28: a machine-checked proof written with AI

For what a proof checker certifies and what it does not (Units 10, 17 and 18). Read in full on 2026-10-28, for what it reports about its own method and its account of other proofs in its field; neither was checked by this course.

[L18] Ben-Shahar, M., Morag, J., Englund, D., and Sulimany, K. (2026). "Quantum Cryptography and Autoformalization in Lean." Preprint, arXiv 2610.17286 (v1 12 October 2026). https://arxiv.org/abs/2610.17286 (accessed 2026-10-31; the abstract record only, still v1). Opened 2026-11-03: the v1 PDF; the sections named below and the quotations were checked against it. Supports, from its own account: the formalisation was developed "through an iterative collaboration between researchers and LLM agents"; agents based on Claude and GPT translated statements into the Lean proof assistant, and the Lean kernel checked each proof against its statement, so the authors' review focused on whether the statement, its hypotheses and its definitions express the intended claim (Sec. II, Fig. 1); that review caught "an earlier AI-generated protocol" that "computed the key as a hash of Alice's and Bob's key strings combined, although neither party had access to both", which led them to a protocol language that enforces locality by construction (Secs. II-III); under the security definition alone "a protocol that always aborts will appear secure", so it is coupled with robustness, a bound on the abort probability, following Ref. [5] (Secs. III.B, IV.C); the protocol assumes an authenticated public classical channel, stated openly (Secs. I, IV.A); the theorems use only Lean's standard axioms (Sec. IV); formal security proofs of optical QKD with decoy states "would additionally require suitable source, loss, and detector models" (Sec. VI); security proofs in the field "can contain gaps, often between individually valid theorems integrated under mismatched assumptions" (abstract); formalising made explicit conditions that two cited results relied on without stating, and the formal statements assume or add them (Sec. VI). May NOT say: that quantum key distribution, or any device, is secure; anything about the correctness of its proof or its quantum-information results, which this course has not checked; that it is peer reviewed (it is a preprint); that the proof was written mainly by AI; that the authors devised the robustness condition; that the paper hides or overlooks the authenticated-channel assumption; that its protocol language rules out every unintended protocol.

Added 2026-10-31: understanding, intelligence and consciousness

For Unit 17's account of three words without agreed definitions, and Unit 18's note on the systems reply. Opened 2026-10-31; what was read is stated per entry.

[T17] Legg, S. and Hutter, M. (2007). "A Collection of Definitions of Intelligence." Technical report IDSIA-07-07 (15 June 2007); Frontiers in Artificial Intelligence and Applications 157: 17-24. arXiv 0706.3639. https://arxiv.org/abs/0706.3639 (accessed 2026-11-03). Opened (the arXiv technical report, 12 pages; the published pages were not seen, so it is cited by section): sec. 1 has "no standard definition", "70 odd" and "a single general and encompassing definition"; secs. 2-4 hold the definitions (Boring's is sec. 3, no. 5); sec. 5 has the common features, the adopted definition and "difficult to argue". Supports: "there is still no standard definition of intelligence"; the collection of "70 odd" informal definitions (18 from dictionaries, encyclopaedias and groups, 35 from psychologists, 18 from AI researchers), which the authors say cannot be complete; examples quoted from it, such as Boring's "Intelligence is what is measured by intelligence tests"; the authors' view that this pessimism is "too strong", and that "a single general and encompassing definition for arbitrary systems was possible", because many definitions share features (a property of an agent interacting with an environment, tied to success at a goal, depending on adaptation to different goals and environments); the definition they adopt, "Intelligence measures an agent's ability to achieve goals in a wide range of environments"; their remark that it is "difficult to argue that there is an objective sense in which one definition could be considered to be the correct one". May NOT say: that the authors think intelligence cannot be defined; that their definition is agreed by others.

[T18] Chollet, F. (2019). "On the Measure of Intelligence." Preprint, arXiv 1911.01547 (v1 5 November 2019, v2 25 November 2019). https://arxiv.org/abs/1911.01547 (accessed 2026-11-03). Opened (v2 PDF): the abstract and sec. II.1.1 argue that skill alone falls short, that priors or data "buy" skill, and that a hard-coded solution shows "the engineer's" intelligence; sec. II.2.1 gives the definition and the "one true" disclaimer; sec. I.2 has "no existing scientific consensus"; sec. II.2.3 asks the questions of a system; sec. III.2 has "a work in progress" and "Test validity is not established". Supports: there is "no existing scientific consensus around any single definition"; two conceptions found in definitions, task-specific skill and general ability to acquire new skills; the argument that "solely measuring skill at any given task falls short of measuring intelligence", because unlimited priors or training data can "buy" skill while masking a system's own generalisation power; a hard-coded chatbot or puzzle solver shows its engineer's intelligence, not its own; his definition, "The intelligence of a system is a measure of its skill-acquisition efficiency over a scope of tasks, with respect to priors, experience, and generalization difficulty", which he says is not the "one true" definition and is not "meant to achieve broad consensus", but meant to be actionable; the questions it asks of any system (its scope, among others); the ARC benchmark, which he calls "a work in progress, not a definitive solution", with test validity "not established". May NOT say: his 2019 assessment of deep learning as a claim about current models; that ARC measures intelligence; that his definition is agreed.

[T19] Nagel, T. (1974). "What Is It Like to Be a Bat?" The Philosophical Review 83 (4): 435-450. DOI 10.2307/2183914. Record from the SEP entry [T22] (which gives the pages as 435-456) and the JSTOR record. Opened 2026-11-03: an unofficial web transcription with neither page numbers nor sections, https://www.philosopher.eu/others-writings/nagel-what-is-it-like-to-be-a-bat/ (accessed 2026-11-03), whose paragraph beginning "Conscious experience is a widespread phenomenon" says "an organism has conscious experience at all means, basically, that there is something it is like to be that organism"; it is cited without a locator. [T22] sec. 2.1 states the same criterion. Supports: the criterion that a being is conscious just if there is "something that it is like" to be that creature. May NOT say: anything further about Nagel's argument.

[T20] Chalmers, D. J. (1995). "Facing Up to the Problem of Consciousness." Journal of Consciousness Studies 2 (3): 200-219. Record from the SEP entry [T22] and the CogPrints record. Opened 2026-11-03: the author's online text, https://consc.net/papers/facing.html (accessed 2026-11-03), without the journal's page numbers, so cited by its sections: sec. 2, "The easy problems and the hard problem", calls the problem of experience "The really hard problem of consciousness" and asks "Why should physical processing give rise to a rich inner life at all?". [T22] sec. 5.4 agrees. Supports: the "hard problem", giving an intelligible account of how phenomenal consciousness could arise from physical or neural processes. May NOT say: Chalmers's own proposed theory, or his list of the easy problems.

[T21] Cole, D. (2004, rev. 2024). "The Chinese Room Argument." Stanford Encyclopedia of Philosophy. https://plato.stanford.edu/entries/chinese-room/ (first published 19 March 2004, substantive revision 23 October 2024; accessed 2026-11-03). Opened in that revision: "brains are machines, and brains think" is sec. 3; the systems reply, Searle's answer and "multiple personality" are sec. 4.1; intentionality and Gardiner are sec. 5.2 (the "owes us" words are Cole's report of Gardiner, not a quotation of him); "no consensus" is the unnumbered "Conclusion". Supports: the systems reply (the man is a part of a larger system; the man does not understand Chinese but the system as a whole does); Searle's answer (memorise the rules); replies to that answer, that the result would be "much more like a case of multiple personality" than identity of Searle with the system (Cole, Block) and the virtual mind reply; that the argument does not "purport to show that no machine can think" ("Searle says that brains are machines, and brains think"); "Searle sees intentionality as a causal power of the brain, uniquely produced by biological processes"; Gardiner's view that "Searle owes us a more precise account of intentionality"; "there is still no consensus as to whether the argument is sound". May NOT say: that the critics are right, or anything about large language models (its section 4.6 discusses them; the course has not used that section).

[T22] Van Gulick, R. (2004, rev. 2014). "Consciousness." Stanford Encyclopedia of Philosophy. https://plato.stanford.edu/entries/consciousness/ (first published 18 June 2004, substantive revision 14 January 2014; accessed 2026-11-03). Opened in that revision: Nagel's criterion is sec. 2.1, access consciousness sec. 2.2, "no privileged or canonical meaning" sec. 2.3, the hard problem sec. 5.4, integrated information in degrees sec. 8.1, functionalism and Dennett's modest eliminativism sec. 8.2, higher-order theories sec. 9.1, global workspace sec. 9.5, integrated information theory sec. 9.6, neural theories sec. 9.7, no single perspective sec. 10. Supports: "conscious" and "consciousness" are used "in a wide range of ways with no privileged or canonical meaning"; Nagel's criterion [T19]; the hard problem [T20]; families of theory: functionalist (a state is conscious in virtue of its functional role), global workspace (information available to a global workspace is conscious at least in the access sense), higher-order (a conscious state in terms of meta-mental self-awareness), integrated information (Tononi: consciousness identified with integrated information, in degrees), and neural theories; eliminativist views that deny consciousness, and "more modest eliminativists, like Dennett"; that no single perspective is likely to suffice. May NOT say: anything about AI systems, or developments after 2014 (the entry does not discuss illusionism; see [T23]).

[T23] Frankish, K. (2016). "Illusionism as a Theory of Consciousness." Journal of Consciousness Studies 23 (11-12): 11-39; reprinted as the target paper of K. Frankish (ed.), Illusionism as a Theory of Consciousness, Imprint Academic, 2017. Pages from the author's own reference list in Frankish, K., "What is illusionism?", Klesis 55, https://www.revue-klesis.org/pdf/klesis-55-04-keith-frankish-what-is-illusionism.pdf (accessed 2026-10-31); the publisher's page refused automated access. Opened: the author's page for the volume, https://www.keithfrankish.com/illusionism-as-a-theory-of-consciousness/. Opened 2026-11-03: the author's accepted manuscript, version of 16 October 2026, without the journal's page numbers, so cited by section, https://keithfrankish.github.io/articles/Frankish_2016_Illusionism_as_a_theory_of_consciousness_eprint.pdf (accessed 2026-11-03). Sec. 1.1: "Illusionists deny that experiences have phenomenal properties", introspective awareness leading us "to misrepresent the states as having phenomenal properties"; sec. 1.6: illusionists "do indeed deny that such things exist" (experiences with phenomenal properties), but "do not deny their existence" (experiences of the kind philosophers so describe), nor "a form of consciousness"; the introduction has "often dismissed out of hand". The words "everyday sense" are the Klesis essay's (sec. 3), not the 2016 paper's. The volume page carries the first two quotations below. Supports: "Illusionism is the view that phenomenal consciousness (in the philosophers' sense) is an introspective illusion": introspection "misrepresents experiences as having phenomenal properties"; the view is "often dismissed out of hand for failing to take consciousness seriously". From the Klesis essay's abstract (read): illusionism is "the view that consciousness does not involve awareness of special 'phenomenal' properties and that belief in such properties is due to an introspective illusion"; and from its sections 2-3: "Illusionists deny that such properties exist"; "I did not say that consciousness itself is illusory, only that phenomenal properties are"; illusionists "can affirm the reality of consciousness", as "having experiences, understood in an everyday sense". May NOT say: that illusionists deny that anyone is conscious in every sense, or anything about AI.

Added 2026-10-31: the incompleteness theorems

For Unit 18's statement of Gödel's second theorem and its conditions.

[T24] Raatikainen, P. (2013, rev. 2025). "Gödel's Incompleteness Theorems." Stanford Encyclopedia of Philosophy, E. N. Zalta and U. Nodelman (eds.). https://plato.stanford.edu/entries/goedel-incompleteness/ (accessed 2026-11-03). First published 11 November 2013, substantive revision 8 October 2025; opened in that revision, and each section given below checked against it. Supports: the second incompleteness theorem, "For any consistent system F within which a certain amount of elementary arithmetic can be carried out, the consistency of F cannot be proved in F itself", F needing "a little bit more arithmetic" than the first theorem does; that it is "a theorem about formal provability" in F, and "does not say anything about whether" F's consistency can be shown "by a conclusive argument, or by a proof generally acceptable for mathematicians. For many theories, this is perfectly possible"; that a formal system's axioms must be "finite or at least decidable"; what consistent and complete mean (all Sec. 1.1); what omega-consistency means (Sec. 2.1) and Rosser's 1936 removal of it (Sec. 2.5); that the version Gödel sent for publication in 1931 "now contained also a statement of the second incompleteness theorem" (Sec. 5); that a complete theory is decidable, and every theory containing Robinson arithmetic is incomplete and undecidable (Sec. 4.2); that the set of sentences provable in arithmetic can be defined in its language and the set of true arithmetical sentences cannot, so "the two cannot coincide" (Sec. 5). Unit 18's sentence that the true arithmetical sentences form a consistent, complete set no procedure can list follows from those statements; the step is the course's, not the entry's wording. Also: applications to the philosophy of mind are "more controversial", and of the Gödelian arguments against mechanism "there is wide consensus that they fail" (Sec. 6.3). May NOT say: anything about language models, which the entry does not discuss; that the second theorem shows a system's consistency cannot be established at all.

Added 2026-10-31: sustainability and industry practice

General starting points for what this course does not teach (Unit 17): the sustainability of AI beyond Unit 4's resource section, and how organisations govern and deploy AI. Each was opened on 2026-10-31; what was read is stated per entry. None is cited for a figure about one query or one model.

[L19] International Energy Agency (2025). Energy and AI. IEA special report, Paris: IEA. Licence CC BY 4.0. https://www.iea.org/reports/energy-and-ai (accessed 2026-11-03). Opened: the report page (published 10 April 2025) has "no AI without energy"; its "Executive summary" page has the other two quotations below. Cited only as a starting point, so without a locator. Supports: "there is no AI without energy"; "There are large uncertainties in the outlook for AI-related electricity demand"; "AI can be a tool in reducing emissions, but it is not a silver bullet and does not remove the need for proactive policy". May NOT say: what one query or one model costs in energy; its 2030 data-centre projection as a forecast rather than a scenario under stated assumptions; anything after its publication.

[L20] Luccioni, A. S., Jernite, Y., and Strubell, E. (2024). "Power Hungry Processing: Watts Driving the Cost of AI Deployment?" Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT '24). DOI 10.1145/3630106.3658542. arXiv 2311.16863. https://arxiv.org/abs/2311.16863 (accessed 2026-11-03). Opened: the arXiv PDF (v3, 15 October 2024); both quotations below are in its abstract. Cited only as a starting point, so without a locator. Supports: deployment cost measured "as the amount of energy and carbon required to perform 1,000 inferences on representative benchmark dataset"; "multi-purpose, generative architectures are orders of magnitude more expensive than task-specific systems for a variety of tasks". May NOT say: what any commercial assistant costs per query; that its figures hold for later models; a carbon figure without its conditions.

[L21] OECD (2022). "Measuring the environmental impacts of artificial intelligence compute and applications: The AI footprint." OECD Digital Economy Papers, 15 November 2022. DOI 10.1787/7babf571-en. https://oecd.ai/en/ai-publications/ai-footprint (accessed 2026-11-03). Opened: the publication page, whose abstract carries the quotations below; the paper itself was not read. Cited only as a starting point, so without a locator. Supports: "Artificial intelligence (AI) systems can use massive computational resources, raising sustainability concerns"; it "distinguishes between the direct environmental impacts of developing, using and disposing of AI systems" and indirect effects of AI applications; it "recommends the establishment of measurement standards". May NOT say: any current figure (it predates the spread of generative AI); that the standards it recommends now exist.

[L22] National Institute of Standards and Technology (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1, January 2023. DOI 10.6028/NIST.AI.100-1. https://www.nist.gov/itl/ai-risk-management-framework (accessed 2026-11-03). Opened: the framework page ("The AI RMF 1.0 is being revised") and the PDF https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf, which has both quotations below. Cited only as a starting point, so without a locator. Supports: "The Framework is intended to be voluntary, rights-preserving, non-sector-specific, and use-case agnostic"; its Core of four functions, "GOVERN, MAP, MEASURE, and MANAGE"; its characteristics of trustworthy AI. The page states that version 1.0 "is being revised"; cite it as 1.0. May NOT say: that it is law or binding; that an organisation citing it follows it.

[L23] Autio, C., Schwartz, R., Dunietz, J., Jain, S., Stanley, M., Tabassi, E., Hall, P., and Roberts, K. (2024). Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. NIST AI 600-1, July 2024. DOI 10.6028/NIST.AI.600-1. https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf (accessed 2026-11-03). Opened: the PDF (64 pages), which has the three quotations below. Cited only as a starting point, so without a locator. Supports: it "defines risks that are novel to or exacerbated by the use of GAI", among them confabulation, "The production of confidently stated but erroneous or false content"; "Currently there is no agreed upon method to estimate environmental impacts from GAI". May NOT say: its training-emissions comparison as a fact about current models; that it is mandatory; anything about its policy status, which was not checked.

[L24] OECD (2019, updated 2024). OECD AI Principles (the OECD Recommendation on Artificial Intelligence, OECD/LEGAL/0449). https://oecd.ai/en/ai-principles (accessed 2026-11-03). Opened: the principles page, which has both quotations below and lists the five principles under "Values-based principles". Cited only as a starting point, so without a locator. Supports: "The OECD Recommendation on AI is the first intergovernmental standard on AI"; "initially adopted in 2019 and updated in May 2024"; its five values-based principles (inclusive growth, sustainable development and well-being; human rights and democratic values, including fairness and privacy; transparency and explainability; robustness, security and safety; accountability). May NOT say: that it is binding; what the 2024 update changed; a number of adherents as current.

Added 2026-11-02: a religious account of what lies outside

A further view beside Unit 18's "no view from nowhere". Opened on 2026-11-02.

[T25] Leo XIV (2026). Magnifica Humanitas. Encyclical letter, 15 May 2026. https://www.vatican.va/content/leo-xiv/en/encyclicals/documents/20260515-magnifica-humanitas.html (accessed 2026-11-03). Opened: the English HTML page, saved and checked character for character (dated "Given in Rome, at Saint Peter's, on 15 May, in the year 2026"); each quotation below is verbatim in the numbered paragraph given, and par. 127 also says that faith's "openness toward the 'beyond'" "originates as a gift from God". Supports: par. 10 warns against the idea that a "language — even a digital one — can translate everything, including the mystery of the person, into data and performance"; par. 127, "it is God himself who overcomes the 'infinite' disproportion"; par. 128, "A person's future is not calculable". May NOT say: that it speaks of a standpoint, of formal systems or of Tarski (the link to Unit 18's standpoint is this course's own reading); that the course endorses a religious claim; that it settles anything about what systems can know.

Added 2026-11-02: accessibility

The standard Unit 15 names for its accessibility checks, and the requirement behind alt text. Opened on 2026-11-02.

[A1] W3C (2023, updated 2024). Web Content Accessibility Guidelines (WCAG) 2.2. W3C Recommendation, current version 12 December 2024. https://www.w3.org/TR/WCAG22/ (accessed 2026-11-03). Opened: the Recommendation of 12 December 2024; the three levels are in the Introduction under "WCAG 2 Layers of Guidance", and the text-alternative requirement and its exceptions are Success Criterion 1.1.1, cited as SC 1.1.1. Supports: three levels of conformance, "A (lowest), AA, and AAA (highest)"; Success Criterion 1.1.1 Non-text Content (Level A): "All non-text content that is presented to the user has a text alternative that serves the equivalent purpose", with exceptions including decorative content, which is to be hidden from assistive technology. May NOT say: that passing automated checks means conformance; that any page of this course conforms; anything about a particular screen reader's behaviour.

Added 2026-11-03: high-risk uses in law

[R5] Regulation (EU) 2024/1689 (Artificial Intelligence Act). Official Journal of the European Union, L series, 12 July 2024. http://data.europa.eu/eli/reg/2024/1689/oj (accessed 2026-11-03; official PDF). Annex III, "High-risk AI systems referred to in Article 6(2)" (pp. 127-129), lists the areas: biometrics; critical infrastructure (safety components); education; employment; access to essential services and benefits, including eligibility for healthcare services, creditworthiness, and life and health insurance; law enforcement; migration, asylum and border control; and the administration of justice. Supports: health, law, finance, safety, immigration and employment are each among the areas EU law treats as high-risk for AI systems that decide about natural persons (Unit 12). May NOT: that the Act applies outside the EU or to a learner's own use of an assistant; that a use in these areas is unlawful; how often decisions are made in these areas.

Added 2026-11-03: what people leave out of a request

[L25] Mondal, S., Bappon, S. D., and Roy, C. K. (2024). "Enhancing User Interaction in ChatGPT: Characterizing and Consolidating Multiple Prompts for Issue Resolution." 21st International Conference on Mining Software Repositories (MSR '24), Mining Challenge, Lisbon. DOI 10.1145/3643991.3645085; arXiv 2402.04568 (accessed 2026-11-03). The authors coded 85 conversations needing more than one prompt (686 prompts) from the DevGPT set of ChatGPT conversations linked from GitHub issues, 2023. Table 2 defines "Missing Specifications": "Prompts miss explicit information (e.g., programming language, operating system, test cases)", 34 conversations; sec. 3.1 calls it, at 40%, one of "the most frequently observed gaps". Supports: in that sample, missing specifications were the most frequent reason a conversation needed another prompt (Unit 7). May NOT: how often people write one-sentence or vague prompts overall (only multi-prompt conversations were coded); anything about students, the general public, other assistants or current models. The coding was by the authors, with no agreement statistic reported.

Added 2026-11-03: quality in reflexive thematic analysis

[E23] Braun, V. and Clarke, V. (2021). "Evaluating and reviewing (reflexive) TA research: A guide for editors and reviewers." Thematic Analysis (the authors' site, University of Auckland). https://www.thematicanalysis.net/editor-checklist/ (accessed 2026-11-03). The twenty questions the authors published in their 2021 paper on quality in thematic analysis (Qualitative Research in Psychology, DOI 10.1080/14780887.2020.1769238, not opened: publisher and repository refused automated access). Question 12 names as "conceptual and procedural confusion" claiming reflexive TA while using "a codebook or coding frame, multiple independent coders and consensus coding, inter-rater reliability measures"; question 7 lists assuming that grounded-theory concepts such as saturation apply to TA. The page also points to their "15-point checklist for quality (reflexive) TA" in [E9]. Supports: reflexive TA has its own quality criteria, which reject those procedures (evaluation guidance). May NOT: that other kinds of thematic analysis (coding reliability, codebook) reject them.

Citing in course materials

In-text, cite by bracket label: (Lovelace 1843, Note G) for [P2], [S1] for a technical point. Where the part of a source that supports the claim is known and has been opened, name it: [C1, p. 303], [S7, sec. 5.2], [T25, par. 10]. Every citation must resolve to an entry on this page. If it does not, it is a defect.