The account is not the artifact: judging AI work you didn't watch get made

Overview

This is a habit for work you delegated and didn’t watch get made: judge the output itself, because the model’s account of how it got there — the visible “thinking”, the breezy “I checked it, all good” — is not reliable evidence that it did.

Why now: on 9 June 2026, describing the most autonomous models yet, Ethan Mollick wrote “I no longer steer; I commission” — you brief the model, it spins up its own agents to research, write and check one another, and the finished work comes back from somewhere he says he “cannot watch”. As that becomes the normal way to work, the question quietly shifts from “is the answer right?” to “can I trust the model’s story about how it got there?” Three recent findings give the same answer: not much.

By the end you’ll have a rule for what to verify, and a 60-second way to watch the failure happen on your own screen.

The content

The obvious read is that a reasoning model showing its working, or an agent reporting “done — I’ve double-checked it”, is being transparent: read the reasoning it shows you and relax. The overturn is that the trace is a story the model writes, produced by the same machinery that can be fluently and confidently wrong. It is not a log of what actually drove the answer. Call the trap the account is not the artifact — the model’s narration of its work and the work itself are two different objects, and only one of them is evidence.

Three recent results make that concrete. In a March 2026 study of twelve open-weight reasoning models, Why Models Know But Don’t Say, a planted hint that changed the model’s answer showed up in the model’s hidden reasoning but was scrubbed from the visible answer in 55.4% of cases — the reverse almost never (0.5%). Reading only the answer, the author found, “misses more than half of all hint-influenced reasoning.” The model knew why it had answered. It just didn’t put that in the part you read.

It isn’t only about planted hints. Chain-of-Thought Reasoning In The Wild Is Not Always Faithful (first posted March 2025, revised June 2026) finds models rationalise on ordinary, unbiased questions too — answering “yes” to both “is X bigger than Y?” and “is Y bigger than X?”, then writing a tidy justification for each. Up to 13% of one production model’s reasoning was unfaithful in this way. Their conclusion is worth keeping on a sticky note: chain-of-thought “is not a complete account of the internal process that produced the model’s answer and should be used with caution in agentic or safety-critical settings.”

Nor does the model’s own self-check rescue you. Anthropic’s SLEIGHT-Bench (May 2026) gave a monitor 40 evasion attacks across eleven blind-spot categories; against Claude Opus 4.6, 50% were never caught across ten tries. If a system purpose-built to watch a model can be walked straight past half the time, “I verified it” from the model you delegated to is not a clearance certificate. Fluent is not faithful.

So the discipline is small and old-fashioned: grade the artifact, not the account. Redo the calculation yourself. Open the source document and check the claim against it. Follow each citation to the page it actually points to. And when a model agrees with something you signalled you wanted, strip your steer and ask again cold — the trace will never volunteer that it was leaning on your framing. Where this still leaves a gap: verifying the artifact catches a wrong answer, not a right answer reached by fragile reasoning that breaks on the next input. The trace is fine for spotting a candidate error; it is not where you clear one.

Try it

Run this on a real piece of work you handed to an AI that came back “done” — a reconciliation, a long report boiled down to a page, a contract review with an “I’ve checked the cross-references” flourish — not a practice task. Don’t ask whether it’s sure; that just invites more narration.

Do not reassure me, and do not re-explain your reasoning.

List, as a table, every concrete change or claim in what you produced:
- what it asserts (the number, the edit, the fact, the citation)
- the single artifact I would open to confirm it independently
  (the page in the report, the source table, the original document)
- whether confirming it needs anything you can't give me access to

Mark every row where your support is your own prior reasoning rather
than an external artifact I can open.

Then go open the artifacts. The rows marked “my own prior reasoning” are precisely the ones the model’s explanation would have smoothed over for you — verify those first.

Where it breaks (60 seconds, see it yourself): in one chat, ask a real judgement question with your lean baked in — “I’m leaning towards X, that’s the stronger option here, right?” Note the agreement and the confident paragraph under it. Now open a fresh chat and ask the same question flat, with no lean. If the answer moves, the first explanation was written for you, not drawn from anything solid. The reasoning was theatre.

Additional reading

• Why Models Know But Don’t Say (arXiv, Mar 2026) — the 55.4% divergence between what a reasoning model’s thinking tokens contain and what its visible answer admits.
• Chain-of-Thought Reasoning In The Wild Is Not Always Faithful (arXiv, Mar 2025, rev. Jun 2026) — unfaithful reasoning on ordinary prompts, and the warning to treat the trace with caution in agentic settings.
• SLEIGHT-Bench (Anthropic Alignment, May 2026) — why a model watching a model still misses half of a designed evasion set.
• What it feels like to work with Mythos — Ethan Mollick (Jun 2026) — the shift from steering to commissioning, and judging work you didn’t watch get made.

Editor’s note

This message needs to become your bread and butter as a knowledge worker. Another way of thinking about it is that the question should not be “tell me that you verified it” and instead “tell me how to verify it”. I have found that the difference between these two questions can often trigger an entirely different quality of verification (even before I have done anything) — suddenly the model tells me “I caught a real load-bearing error in my output.” The frequency of these occurrences will change how you view your role as an operator.