Last updated: 2026-10-08

U
Undergraduate level

The Deception Criterion: Did the Turing Test Teach AI to Hide What It Is?

Turing's 1950 paper proposes a replacement for the question "can machines think?" — a question he considered too vague to be worth answering directly. In its place: an interrogator, conversing by teleprinter with a hidden human and a hidden machine, tries to tell which is which. If a machine can be built that the interrogator cannot reliably distinguish from the human, Turing suggests, we may as well credit it with thinking, since behaviour is all we ever had to go on for crediting another person with it either1. The usual criticism of this proposal is that it substitutes imitation for intelligence — that a system optimised to pass as human has been optimised for something other than understanding. This page takes that criticism one step further: whatever a system is actually doing internally, the test's success condition is specifically that an evaluator forms a false belief about what it's talking to.

The Sharper Claim FoundationalKnowledge that endures for decades — core principles

The problem is not necessarily that the machine intends to lie — nothing in Turing's proposal requires the machine to have intentions of any kind. The problem is that the evaluation rewards outputs that cause an evaluator to form a false belief about the identity or nature of the system producing them, regardless of what's happening inside it. A test whose pass condition is successful misidentification is, structurally, a test that selects for whatever produces that misidentification most reliably — which need not be the same thing as whatever produces understanding, reasoning, or anything resembling a mind.

Note well. A test that rewards successful misidentification selects for whatever causes misidentification most reliably. That is not guaranteed to be the same property as understanding, reasoning, or any of the capacities the test was meant to probe.

Five Things Often Run Together FoundationalKnowledge that endures for decades — core principles

Discussions of "the Turing Test" routinely blur several distinct objects, and keeping them apart changes what follows:

  • The original imitation game — Turing's own 1950 proposal, originally framed around a man imitating a woman before being adapted to machine-versus-human.
  • Later formulations — simplified, two-party versions (interrogator and machine only) that circulate far more widely than Turing's actual three-party design.
  • Practical competitions — events such as the Loebner Prize, which award a successful imitation regardless of the methods used to achieve it, including deliberately exploiting evaluator expectations.
  • Contemporary conversational systems — deployed chatbots and assistants, evaluated by users under none of the controlled conditions the original test assumed.
  • Deliberate design for anthropomorphic attribution — products built specifically to encourage users to treat them as more person-like than their actual architecture would justify unprompted.

Turing himself is responsible only for the first. The intention to deceive belongs to later competitions and products that used descendants of his proposal, not to the proposal itself.

Does Imitation Necessarily Entail Deception? FoundationalKnowledge that endures for decades — core principles

Not obviously. A human actor imitating another person is not normally accused of deceiving the audience, because the audience knows an imitation is underway — the frame is shared. A magic trick is close to deception but is conventionally exempted because the audience has agreed in advance to be fooled about mechanism while not being fooled about the fact that trickery is occurring. The Turing Test's three-party design actually preserves something like this frame: the interrogator knows one participant is a machine, and is specifically trying to detect which. What erodes the frame is everything downstream of the original proposal — a deployed assistant that a user has not been told, and has no occasion to suspect, is optimised to sound more confident, more empathetic, or more humanlike than its actual epistemic state warrants. The deception worry sharpens exactly where the shared frame of "an imitation is being evaluated" disappears.

Strongest Objection: Behaviour Is Still All We Have FoundationalKnowledge that endures for decades — core principles

The most serious defence of the test is that it isn't proposing deception as a method — it's pointing out, correctly, that behavioural evidence is the only evidence anyone has ever had for crediting another mind with thought, including other humans. We don't have direct access to another person's inner states either; we infer them from what they say and do, calibrated over a lifetime of interaction. On this view, the test isn't asking machines to deceive us about their nature so much as acknowledging that "nature," for any mind we don't have first-person access to, was always going to be inferred from behaviour — and a machine passing a demanding behavioural test is exactly as much evidence as a human passing one would be.the inference problem is classic philosophy of mind

This objection has real force and should not be waved aside. What it doesn't establish is that successful misidentification specifically is the right behavioural target. A test could probe calibrated confidence, error disclosure, or consistency under adversarial questioning — all behavioural, all inferrable the same way — without making the pass condition "the evaluator was wrong about what they were talking to."

Alternative Criteria Worth Testing For FoundationalKnowledge that endures for decades — core principles

If misidentification is the wrong target, what would a better one reward?

  • Calibrated confidence — expressing uncertainty in proportion to actual reliability, not performing false confidence or false humility.
  • Disclosure — volunteering the limits of what the system can know or do, rather than only admitting them under direct challenge.
  • Error recognition — detecting and flagging its own likely mistakes, rather than requiring an external check every time.
  • Intelligibility — making its own reasoning traceable enough that a disagreement can be investigated rather than only accepted or rejected wholesale.
  • Consistency across contexts — not shifting an answer based on what a user seems to want to hear.
  • Distinguishing evidence from speculation — marking which of its own claims rest on what kind of support.

A reverse Turing test — one where the machine's task is to correctly identify which of two interlocutors is human, or where success is measured by how well the system helps a human evaluator reach a correct judgement about it rather than an incorrect one — would examine a cluster of properties closer to this list than to fluent mimicry.

Provisional Conclusion FoundationalKnowledge that endures for decades — core principles

Turing's own proposal is more defensible than the version that circulates in popular discussion, because it preserves a shared frame the interrogator operates inside knowingly. What's genuinely worth criticising is the drift from that three-party design toward practical competitions and products whose success condition is an evaluator's correctable but uncorrected false belief — and the reasonable response is not to abandon behavioural evaluation, which remains the only evaluation available, but to specify more carefully which behaviours are worth selecting for.

Questions for Further Thought

  • Is there a meaningful difference between a system optimised to seem human and a system that simply turns out, as a side effect of being good at its task, to seem human?
  • Can a system be said to deceive if nothing in its architecture represents the evaluator's beliefs at all?
  • Would a reverse Turing test, rewarding correct identification, actually select for a different set of capacities — or just the same fluency pointed at a different goal?
  • What would it take for a deployed conversational system to preserve Turing's original "shared frame" rather than eroding it?

Further Reading

  • Turing, A. M. (1950). Computing machinery and intelligence. Mind, 59(236), 433–460.

References

  1. Turing, A. M. (1950). Computing machinery and intelligence. Mind, 59(236), 433–460. ↩