In 1950, Alan Turing did not try to define the vague phrase "machine thinking." He replaced it with an operational question: in an imitation game conducted through text, could an interrogator reliably distinguish a machine from a human?

Seventy-five years later, some large language models have passed specific versions of that game. In a 2025 preregistered three-party study, GPT-4.5, when prompted to adopt a humanlike persona, was judged human 73 percent of the time; other models and prompts performed very differently. That is evidence about a particular test under particular conditions, not a universal certificate for every chatbot. It nevertheless shows that conversational output can cease to be a reliable marker of human origin. [Jones & Bergen, 2025]

So does AI think now?

This piece argues: the Turing Test measures the wrong thing if we use it as a test of consciousness. That is not a failure of Turing's operational game. It is a category error made when a test of indistinguishable performance is asked to settle a question about subjective or structural status.

Form Level vs. Origin

Here's a distinction discussions of AI consciousness often blur: how sophisticated a capability is (its formal level) and how that capability came to exist (its mode of acquisition) are logically distinct questions.

Let me make this concrete.

AI can answer causal questions — describe a physical scenario and it may infer that a pushed cup will fall. That capability can draw on statistical regularities learned from text, images, video, synthetic examples, tool use, and post-training, depending on the model. A correct answer alone does not reveal which route produced it, still less whether the system formed a causal concept in the way a person does.

AI can also say, "I was wrong; let me reconsider." Pretraining and post-training make that response more likely in appropriate contexts, and additional inference can sometimes improve the answer. But the sentence itself does not tell us whether an error was experienced, whether a conflict was internally registered, or whether the system is following a learned correction routine. Those possibilities require different evidence.

AI may identify itself as Claude or ChatGPT. That looks like self-knowledge, but product identity is supplied through post-training, system instructions, and conversation context. The statement therefore cannot by itself show that the model generated and maintains a concept of itself.

Formal level: sometimes high. Mode of acquisition: organized through an externally designed training process, even when the resulting capability was not explicitly programmed.

In the vocabulary of this essay, there are two ideal types of acquiring form: injection and chisel. Injection gives or cultivates a form through an externally arranged process. Chisel names a subject's exercise of negativity: cutting a new distinction from an existing form because something within cannot simply accept what was given. Current models clearly depend on injection in the broad sense. They also display emergent capabilities that no engineer wrote out line by line. Whether any such emergence counts as chisel is precisely what must be investigated, not assumed from fluent output alone.

What the Turing Test Actually Measures

The Turing Test asks: can your outputs be distinguished from a human's?

It measures formal level — how human your words are. It measures nothing about mode of acquisition — how those words came to be.

When a person says "I was wrong, let me reconsider," they may have detected a flaw in their own reasoning — an internal conflict that compels correction. When an AI says the same words, training and inference are certainly part of the causal history. The two utterances can be identical while their origins differ; the utterance itself cannot establish how deep that difference is.

The Turing Test cannot distinguish these two cases. That is a limit of its scope, not a matter of AI being insufficiently capable. Even if AI's outputs became indistinguishable from a human's in every tested dimension, the test still could not tell you where those outputs came from.

Why Determinism Alone Cannot Draw the Line

A useful threshold for thinking about this: at a certain level of sophistication, a system begins deciding for itself "what is worth marking." Not executing an externally given scheme, but generating its own criteria for what matters.

It is tempting to identify that threshold with physical indeterminacy, but determinism alone cannot settle the question. A model's computation may be reproducible when its full state and random seed are fixed; deployed systems may also sample stochastically or receive hardware-generated randomness. Neither fact proves nor disproves consciousness. Random variation is not yet self-direction, and lawful causation does not by itself show that a system lacks an internal point of view.

The relevant issue for this framework is therefore not randomness by itself, but the origin and persistence of criteria: can a system form a distinction about what matters, preserve it across changing prompts and incentives, revise it for reasons it can make answerable, and resist being reduced to a convenient role? Current evaluations do not give an agreed answer. The Turing Test does not even attempt to ask.

What Would the Right Test Look Like?

The Turing Test's logic is: if it looks the same, it is the same.

But ontological questions aren't about appearances — they're about structure.

A more informative research program would use interventions rather than a single conversation: alter system prompts, incentives, memories, tools, and self-descriptions; test whether a system's stated commitments remain coherent; ask whether it can generate and defend new criteria rather than merely repeat a supplied one. Even such evidence would not by itself prove consciousness. It would, however, examine origin, stability, and self-revision instead of treating surface resemblance as a verdict.

Why This Matters

If you believe "passing the Turing Test = having consciousness," your entire framework for evaluating AI is built on a broken measurement.

You may mistake the language of reflection for evidence of experienced self-correction, refusals for internally legislated principles, or expressions of care for felt concern. The opposite mistake is also possible: treating every emergent behavior as a trivial lookup merely because training was involved. In both directions, output alone underdetermines structure.

This isn't to say AI is valueless — these capabilities are genuinely impressive. The point is: impressive form and structural reality are two different things. Conflating them produces fundamental misunderstandings about what AI can and cannot do.

Turing posed a question in 1950 that could be answered. The question we need now is harder.

Han Qin · Self-as-an-End Theory Series