Unit 18 activity: Alles Lüge! First- and second-order reasoning

Read this after attempting the activity.

This key does not restate the answer to cp18-sort-the-claims, and it does not paraphrase it either. Sorting claims by what supports them is the learning objective, and the checkpoint gives targeted diagnostics if you go astray. course/lab/checkpoints.py is the canonical machine-readable source of checkpoint answers, and this key is not a second copy of one. Answers do appear elsewhere in the materials — the guidance edition of the handbook carries the checkpoint data and its explanations — so the point is that they are generated from that one source rather than restated by hand here. The six statements discussed below are deliberately disjoint from the three the checkpoint uses: no statement here is the checkpoint's, and no statement here is a restatement of one of the checkpoint's.

Scope. This unit teaches established, published material only. There is no research framework in it to mark, and answers that reach for one have gone outside the unit.

Citation shorthand. Bracket labels resolve to course/references.md: [C1] von Foerster, [C2] Mead 1968, [C3] Maturana and Varela, [C4] the Macy Conferences, [C5] Bateson, [C6] Gödel 1931, [C7] Tarski 1936; [G1] to [G4] support the register; [Q1], [Q2] and [Q3] are the three songs. Fuller treatment of this unit's argument: course/second-order.md.

The part that counts is the reasoning, not the label. Placing a statement differently but naming a real citation, a real tradition or a real missing step is doing the exercise. Landing on the intended bucket and writing "it sounds like a theorem" is not.

Task 1 — the six statements

A. Incompleteness

Proved theorem. Gödel 1931 [C6]. Any consistent formal system strong enough to express arithmetic contains true statements it cannot prove.

The tell is that it has a date, a name, and a place in every logic textbook. Marking it "established research programme" because it is unfamiliar confuses unfamiliar with unsettled — the mirror image of the error waiting in statement B, and worth naming as such.

What would unsettle it: an error in the proof, or a change to the hypotheses. Note that the statement is conditional — consistent, sufficiently strong — and almost every misuse of it in the wild starts by dropping one of those conditions.

B. Mind as a property of a system of relationships

Established research programme. Bateson [C5], and the wider tradition [C1], [C3]. Published, developed over decades, widely built on, and still an account rather than a theorem.

The tell is the phrase "is better described as". The statement is not reporting a result; it is proposing that a particular way of describing something is the right one. Descriptions of that shape are argued for, adopted, extended, and sometimes displaced. They are not looked up and settled.

Say plainly in discussion that this category is not a polite way of saying "wrong" or "weak". It is a statement about what kind of support the claim has.

What would settle it either way: sustained independent development that makes the account do work nobody could do without it, or a demonstration that it fails on the cases it was built for. Note that "better described as" invites the question better for what purpose, and asking it is finding the seam.

C. The Macy Conferences

None of the three buckets fits, and that is the point. This is a historical claim, and historical claims are sorted as established, contested, or unsourced — the instrument Unit 16 works through, and named in the activity itself so that it is available whether or not you did that unit first.

Sorted that way it is established — course/references.md [C4] records the series, the dates, the funder and the chair.

The best answers do two things: notice that the bucket scheme does not fit, and reach for the other instrument instead. Doing only the first still counts, because recognising that you are holding the wrong instrument is most of the skill.

The trap is marking it "established research programme" because cybernetics is a research programme. The statement is not about the programme; it is about ten meetings, their dates, their funder and their chair. Those are facts with a record, not a way of framing problems.

D. "Gödel shows a language model cannot verify its own output"

Overclaim, and the most interesting one on the list, because both halves of it are true and the join is not.

Gödel 1931 [C6] is a proved theorem. It is a result about formal systems of a specific kind. A language model is not that object. Carrying the result across is an argument, requiring premises about what corresponds to what, and no such argument appears in the sentence.

The failure here is not falsity. The conclusion might even be defensible on other grounds. The failure is a theorem being used as a licence for a claim about a different kind of system, with the transfer step unstated. That move is common, it reads extremely well, and it is precisely the thing this module has been training you to catch.

"The theorem is proved, the transfer is not, and the sentence hides the seam" is worth much more than "overclaim" with nothing behind it.

E. "An AI system's report about itself carries no information at all"

Overclaim. This is the second overclaim from the student page wearing practical clothing, pushed one step further.

Two separate faults. First, it does not follow: that an observer is inside the system says nothing about whether its reports correlate with anything. Second, it is operationally empty — if a self-report carries nothing, you have no way to decide which outputs deserve a check, which in practice means you check nothing, which is exactly where blanket trust also lands you.

What refutes it from your own work: Unit 8. Generated output that was correct, generated output that was wrong, and a check that told you which was which. You were not applying a presumption; you were comparing against a record.

Describe behaviour, not intent. "The system reported X" is the right form; "the system believed X" is not, and slipping into the second is drifting away from what you can actually evidence.

F. Scope of Gödel and Tarski

Proved theorem, in the sense that matters here: this is a statement about what [C6] and [C7] actually say, and you can check it by reading the theorems' hypotheses. Accept "established" with that reasoning.

The generous alternative reading is that F is a methodological remark rather than a result. Saying so, and then saying the remark is checkable against the theorem statements, understands it better than the bucket alone shows.

What must not happen is marking F an overclaim. F is the caution, not the leap. Reading the caution as itself an overreach inverts the section you most needed.

Task 2 — a claim of your own

Step 2, the index

Typical results, and none of them is a failure:

• Observer — almost always implicit. The pack has one author and it feels redundant to name them. It is not redundant to a reader who has three packs on their desk.
• Context — usually stated, because Unit 17's template forces it. This is the one the module worked hardest on, and it shows.
• Record — often half-stated: "the checker reported no divergence" names the check but not where the output now lives, or whether it can be produced again.

A good answer marks at least one of the three as assumed. Marking all three as stated almost certainly means you read your own sentence generously, which is the one reading a sceptical reader will not give it.

Step 3, the clauses

A well-built pack usually contains a record-supported result: there is a record, the reader can consult or reproduce it, and the sentence states the range, convention, and properties actually checked. Do not call this a "proved theorem" or an "established research programme". Reusing Task 1's instrument on this kind of claim is the same error statement C is there to provoke.

The instructive outcome is finding an overclaim in your own writing. The common form is a scope slip: a check that ran over a specific range, under a specific convention, reported as though it established correctness in general. That is not dishonesty and it does not need punishing. It is the same structure as the Note G table — a narrow supported claim and a broad readable one, and the broad one is easier to write.

Reasoned judgement is usually present beside the record: which discrepancy mattered, which convention was the sensible default, and what counted as resolved. It should be attributed to you by name, not dressed up as something more.

Step 4, the honest sentence

The intended discovery is that the record-backed part of a defence is usually smaller than it felt while writing it, and that the remainder is not worthless — it is judgement, which is exactly the thing the module says must be attributable to a person. Concluding "so my defence is weak" misreads the exercise. The point is to be able to say which part is which.

Common ways this activity goes wrong

Sorting by how the sentence sounds. Statements B, D and F are all written in the register of established results. That is not a trick; it is what most writing about this material looks like, including writing produced by a language model. The sorting has to be done on provenance, not on prose.

Treating "established research programme" as a downgrade. It is not. A programme can be decades old, published, influential and worth taking seriously while still not being a theorem. The unit says so explicitly and the checkpoint tests it.

Reading the unit as a case against AI systems. If you leave saying "so it cannot be trusted", the second overclaim has not been closed properly. Nothing here establishes a base rate for correct output. What it teaches is how to match the strength and independence of a check to the claim and its risk.

Collapsing the whole thing into "everything is relative". The index is the opposite of that. An indexed claim is more checkable than an unindexed one, because it says what would have to be compared with what. The Bernoulli convention is the proof: naming it turns an unresolvable disagreement into a resolved one.

Why this unit is written the way it is

• It comes last, and nothing before it depends on it. No checkpoint, activity or definition in Units 0, 2, 3, 6, 7, 8, 10, 16 and 17 needs this unit; those units carry a pointer to it and nothing more. You can work through the module without reaching here and lose none of the method. This unit is the part that asks what the method cannot reach — which is a different question, and one you are entitled to leave until you want it.
• It raises the issue; it does not settle it. The scope is fixed to established, published material. Where the argument could be extended — however interesting the extension — the unit stops, because past that line it would be doing the thing it warns about.
• Section 4's second half is the unit. "What they do not say" is the point, not a caveat attached to one. If you read only one part of this unit, read that.
• Statement D is the one to sit with. A real theorem, a real system, and an unstated transfer between them is the shape of a large fraction of confident writing about AI and formal limits.
• The atmosphere never replaces the citation. Every claim in this unit keeps its source or its uncertainty label, including the ones on the darkest pages. A unit about unwarranted certainty that asked for your trust on mood would be refuting itself.
• If the status of any claim in course/second-order.md changes, that file is updated first and this unit follows. This unit holds no independent position that could drift out of step with it.