The reversed Turing test: can you prove you think?

Alan Turing proposed his imitation game in 1950. A human interrogator exchanges typed messages with two interlocutors — one human, one machine — and has to decide which is which. If the machine fools the interrogator consistently, Turing argued, it has earned the word “thinks.” The test was never an operational definition; it was a provocation. But it framed the question that animated AI research for seventy years: can a machine perform cognition well enough to be mistaken for a person?

That question is now obsolete. GPT-4, Claude 3, Gemini 2 — they passed Turing’s test somewhere around 2023, in the casual version of the game, without anyone running a formal ceremony. The reasonable question has flipped. The interesting problem in 2026 is not whether the machine can pass for you. It is whether you can still pass for yourself.

This is the reversed Turing test. The interrogator knows one of the two interlocutors is you, and the other is the model that produces your writing, your presentations, your code, your strategic plans. The interrogator can read both outputs, but only one. Their question: can you demonstrate that the thinking behind the work you signed was actually yours?

What the question actually asks

The surface version of the test is a performance check: could you reproduce the argument, the decision, the piece of writing, without the model? Most people who use AI daily find that surface version harder to pass than they expected. The Kosmyna study at MIT (2025) showed that 83% of essay writers could not recall a single sentence from the text they submitted under their name five minutes earlier. A recall failure is a weak form of failing the test.

The deeper version is a defense: asked to steelman the opposing position, identify the three weakest claims in your own argument, predict what would falsify your conclusion — can you? The model can; it was trained to. If you outsourced the thinking, you outsourced the defense too. The difference between the generative and the defensive mode is the clearest diagnostic of where the cognition actually lives.

The deepest version is a judgment check: a novel case, not in the training data, not in the context window, arriving through a channel the model cannot see. A decision has to be made. The model is not available. What do you do? The answer reveals what has been reinstalled in you and what has been left in the machine.

A simple matrix

The table below is the short form of the test. It maps four common deliverables against the three levels of the reversed Turing question. A healthy Cognitive Partisan clears all three levels for any deliverable they ship. A reader with significant cognitive debt will fail level two or three for most deliverables they ship, often without noticing.

Deliverable	Level 1 (Reproduce)	Level 2 (Defend)	Level 3 (Judge)
An essay or memo	Restate the thesis in your own words, a week later, without notes.	Name the three weakest arguments in it.	Re-argue it in 200 words for a hostile audience.
A code change	Explain what it does, line by line, without the diff in front of you.	Describe one failure mode it introduces.	Decide whether to ship it in a production incident.
A hiring recommendation	Reproduce the ranking and the reason for each candidate.	Argue the case for the person you ranked last.	Extend the decision to a different role you didn’t interview for.
A market-strategy doc	Reproduce the three main moves and their sequencing.	Describe the competitor counter-moves.	Decide which move to pull if the market shifts 30% tomorrow.

What the matrix forces you to see is that level 1 — reproduce — is the weakest version of the test and the one most people quietly fail already. Level 3 — judge — is the one that determines whether you are useful to an employer, a client, a reader, a partner. Model-first workflows thin level 3 fastest because the model does not have to actually stand behind any of its outputs. You do.

The stakes are not intellectual

It would be easy to read the reversed Turing test as an academic exercise — a question for philosophers of mind, not for people with calendars. That reading misses the stakes. In a work environment where everyone has access to the same models, the only value an individual contributor brings is the thinking that is indistinguishably theirs. If the thinking happens in the machine, the marginal value of the person approaches the cost of the subscription.

This is the workplace version of Nick Bostrom’s paperclip problem, only inverted. You do not lose your job to a robot that outperforms you. You lose it to a colleague who uses the same robot while still being able to produce the judgment the robot cannot. The reversed Turing test is the operational test for that colleague: can they pass level 3 when the call gets hard? Can you?

This is also why the book’s author refuses the “anti-AI” framing. Anyone who spends a lot of time with models in 2026 has seen what they can do. The argument is not to use them less. It is to insist that the thinking that matters still happens in the place that will be held accountable for it.

Where the circuit lives

Level 1 — reproduce — lives in the Archive layer: the memory store built by handwriting, spaced retrieval, and deep reading. Sparrow, Liu, and Wegner showed in 2011 that people who expect to retrieve a fact fail to encode it in the first place; that is the Archive layer atrophied in miniature. The protocol reinstalls the Archive by interrupting the expectation of retrieval — by insisting that a fact committed to the page by your own hand is the version you will use again.

Level 2 — defend — lives in the Loop layer: the reasoning cycle of holding, testing, steelmanning, rejecting, iterating. Lee and Sarkar at Microsoft Research (2025) measured a shortening of this loop in heavy AI users — users reported both less cognitive effort and less confidence in the result, a combination that is only coherent if the loop has compressed. The protocol reopens the loop by installing structured disagreement: Socratic mode, steelman mode, falsification mode as default prompting patterns, all of which turn the model from a generator into an adversary.

Level 3 — judge — lives in the Voice layer, the output layer, the place where a functional Flesh, Archive, Lens, and Loop produce a sentence or a decision that is unambiguously yours. Voice is the most visible thing in the stack and the one that degrades from the top down: first the sentence loses grain, then the paragraph loses argument, then the decision loses signature, until what is “yours” is only the prompt. Reinstalling Voice requires writing the draft before the model touches it, every time, for thirty days, which is what the protocol does.

The test is private

The reversed Turing test never appears on a dashboard. No manager runs it on you. The interrogator in the thought experiment is you, six months from now, in a meeting with no laptop, asked a question you cannot Google. The work you showed up with in 2026 either has a defender behind it, or it does not. You are the only one who knows which.

This is the shape of the Anti-AI Partisan’s standard. Use the machine harder than most people do. Keep the brain that can answer for the machine’s outputs with no machine in the room. The question that matters is not whether the model can do it — the model can. The question is whether, when the conversation goes past the context window, the thinking continues in you.

FAQ

Is this a real test anyone can administer? There is no formal instrument yet. The matrix above is a short-form self-check, and the Anti-AI Brain Score quiz is the closest available diagnostic for the underlying circuits. The value of the phrase “reversed Turing test” is as a frame, not a rubric — it puts the burden of proof where it now belongs.

Is this the same as the “AI authorship” question? Overlapping but distinct. AI-authorship debates are about attribution, disclosure, and the policy of who gets credit for what. The reversed Turing test is about cognition — whether the thinking that produced the work can still be reconstructed by the person claiming it. You can be transparent about AI use and still fail the reversed test.

Does the test apply to routine tasks? Only loosely. The right question for a routine task is whether automating it frees cognitive budget for a non-routine one. The reversed test applies to the judgment calls that define a person’s professional identity — the work you would not want to discover had been produced by a machine you were not supervising.

Where is this in the book? Chapter 3 introduces the framing formally. Chapter 14 (Part III) extends it into a practice — a weekly exercise the author calls the “defense drill” that substitutes for traditional performance review in teams that ship with AI.

---

Definitions of the terms used above — Flesh, Archive, Lens, Loop, Voice, Cognitive Partisan, cognitive debt — are on the glossary. The studies cited are on the research page. Pre-order The Anti-AI Brain at $0.99 until launch day.