The Turing test is a thought experiment proposed by mathematician Alan Turing in 1950 to sidestep the thorny question “can machines think?” by replacing it with something observable: can a machine convince a human it is also human, purely through conversation? For decades the test lived mostly as a philosophical talking point, but a 2025 study found that GPT-4.5, when given a humanlike persona prompt, was judged to be the human participant 73% of the time, outperforming the actual humans it was tested against. That result has reignited arguments about what, exactly, the test proves and whether passing it tells us anything meaningful about intelligence.
What Turing Actually Proposed
Turing’s original 1950 paper in the journal Mind described what he called the “imitation game.” Three players are involved: two hidden participants and one interrogator. The interrogator communicates with both through text and tries to figure out which is the machine and which is the human. If the machine can fool the interrogator at a rate comparable to (or better than) the human player, it passes.1Oxford Academic. COMPUTING MACHINERY AND INTELLIGENCE – Section: 1. The Imitation Game Turing himself predicted that by the year 2000, machines would be able to fool about 30% of interrogators after five minutes of questioning. That turned out to be optimistic for the technology of 2000, but far too modest for the technology of 2025.
A persistent misunderstanding is that Turing intended the test as a rigorous definition of machine intelligence. He did not frame it that way. He suggested it would be reasonable to regard any machine that could reliably fool interrogators as a thinker, but he never claimed the test was a necessary condition for thought.2Nature Human Behaviour. The Turing test is not a good benchmark for thought in LLMs That distinction matters, because much of the modern debate treats the Turing test as a pass/fail exam for consciousness, which was never its purpose.
The First Controlled Passes
In a pair of preregistered, randomized experiments published in 2025, researchers put four systems through a standard three-party Turing test. Each human participant spent five minutes chatting simultaneously with another real person and with one of the AI systems, then decided which conversation partner was human. The results varied dramatically by system and by setup. GPT-4.5 with a humanlike persona prompt was selected as the human 73% of the time, which was significantly higher than the rate at which interrogators picked the actual human. LLaMa-3.1-405B with the same persona prompt achieved 56%, roughly equal to the real humans. Older or unprompted systems did far worse: ELIZA scored 23% and GPT-4o scored 21%.3PubMed Central. Large language models pass a standard three-party Turing test
The persona prompt turned out to be critical. When GPT-4.5 and LLaMa-3.1 were tested without instructions to behave like a specific kind of person, their pass rates dropped to 38% and 36% respectively, barely outperforming the primitive 1960s chatbot ELIZA.4PubMed Central. Large language models pass a standard three-party Turing test In other words, the raw language ability of these models was not enough on its own. They needed coaching on how to present themselves, adopting a casual tone, using filler words, perhaps even making the occasional typo, to cross the threshold. A third study in the same research program extended conversations to fifteen minutes and found two persona-prompted models still achieved pass rates of 56% and 59%, suggesting the effect was not just a quirk of short interactions.
An earlier and somewhat different online Turing test, conducted with GPT-4, found that its best-performing prompt configuration passed in about half of games, well above ELIZA at 22% and GPT-3.5 at 20%, but still below the human baseline of 66%.5ACL Anthology. Does GPT-4 pass the Turing test? The jump from GPT-4’s roughly 50% to GPT-4.5’s 73% happened in about a year, which gives a sense of how quickly the landscape is shifting.
What Judges Actually Look For
If the Turing test were purely about intellectual ability, you might expect interrogators to quiz the AI on math, logic, or obscure trivia. In practice, that is not what drives their decisions. In the GPT-4 study, participants reported that their judgments were based mainly on linguistic style, which accounted for about 35% of decision rationale, and on socioemotional traits like warmth, humor, or awkwardness, which accounted for another 27%.6ACL Anthology. Does GPT-4 pass the Turing test? Intelligence in a narrow sense, the ability to answer hard questions correctly, was not the main thing people used to tell humans from machines.
This finding is consistent with why persona prompts matter so much. A model that responds in polished, perfectly grammatical sentences reads as robotic to most people, even if every answer is factually correct. A model that occasionally says “lol” or goes on a brief tangent about its weekend feels more human. The test, in practice, measures social plausibility more than cognitive power. That is arguably what Turing intended, but it also means the test rewards performance over substance.
Cultural context adds another layer. A study of a large language model tested in Finnish found that participants relied heavily on colloquial Finnish as a marker of human authorship.7arXiv. Cultural Competence in Context: A Large Language Model Passes the Turing Test in Finland When the model produced natural-sounding slang and informal phrasing, judges were more likely to believe it was human, regardless of the content of its answers. When it defaulted to a more formal register, suspicion grew. This suggests that Turing test results are not culturally portable. A system tuned for casual American English may fail in Finnish or Japanese if it does not match local conversational norms.
Why Humans Are Easy to Fool
People have a well-documented tendency to attribute human qualities to things that are not human. We give names to our cars, feel guilty about neglecting a Tamagotchi, and read emotion into a robot’s tilted head. Researchers call this anthropomorphism: the process of projecting humanlike motivations, emotions, or mental states onto non-human entities.8ScienceDirect. AI anthropomorphism and its effect on users’ self-congruence and self–AI integration: A theoretical framework and research agenda – Section: 3.1. Building block one: self-congruence with anthropomorphised AI agents In a Turing test scenario, this tendency works in the machine’s favor. When a chatbot produces something that sounds even vaguely emotional or personality-driven, the interrogator’s brain is already primed to interpret it as evidence of a real person on the other end.
Research into how users relate to AI agents shows that people go further than simply believing the agent might be human. They begin to experience a sense of personality alignment, feeling that the agent’s “character” matches their own self-concept.9ScienceDirect. AI anthropomorphism and its effect on users’ self-congruence and self–AI integration: A theoretical framework and research agenda – Section: 3.1. Building block one: self-congruence with anthropomorphised AI agents If the AI’s persona comes across as friendly and a bit nerdy, a friendly, nerdy interrogator may feel a rapport that biases them toward believing they are talking to a real person. This is not a flaw in the interrogators so much as a feature of human cognition that the Turing test does not account for. We are social creatures who default to treating conversational partners as people until proven otherwise.
The Deception Problem
The most persistent philosophical objection to the Turing test is that it rewards deception rather than intelligence. A machine that passes is not necessarily thinking; it is successfully pretending to be something it is not. Critics have pointed out for decades that the standard criterion for passing is tied to the AI’s ability to deceive the interrogator about its identity, and there is no obvious reason why intelligence should require the ability to deceive.10SpringerLink (AI & SOCIETY). Intelligence is not deception: from the Turing test to community-based ascriptions
This objection has grown sharper now that large language models can actually pass the test. When the test was purely hypothetical, debating whether deception equals intelligence was an academic exercise. Now that GPT-4.5 can fool three out of four interrogators, the question has practical stakes. If a customer-service bot passes the Turing test, is that a sign of intelligence or a sign that someone has designed a very effective piece of social engineering? The fact that persona prompts, essentially instructions on how to act human, are what push models over the threshold supports the deception reading. The models are not arriving at humanlike behavior through understanding; they are following a script that makes their outputs more socially plausible.
A related critique holds that the Turing test is simply the wrong kind of benchmark for evaluating large language models. One argument, published in Nature Human Behaviour, contends that passing the test tells us more about the limitations of human judgment under conversational pressure than about the capabilities of the machine.11Nature Human Behaviour. The Turing test is not a good benchmark for thought in LLMs A five-minute text chat is a narrow, socially loaded context. Humans are bad at it for reasons that have nothing to do with the machine being smart: they overthink, they anchor on irrelevant cues like typing speed, and they fall prey to the anthropomorphism bias discussed above.
Alternatives That Test Different Things
Frustration with the Turing test has fueled the development of alternative benchmarks designed to measure something closer to genuine reasoning rather than social mimicry. These alternatives share a common goal: evaluating whether an AI system can handle tasks that require understanding, not just pattern-matching.
The Winograd Schema Challenge asks AI systems to resolve sentences that hinge on commonsense knowledge. A classic example: “The trophy does not fit in the brown suitcase because it is too big.” What does “it” refer to? Humans find this trivially easy. For years, machines did not, because the answer depends on understanding physical objects and their properties rather than statistical word patterns.12Proceedings of the AAAI Conference on Artificial Intelligence. The Winograd Schema Challenge: Evaluating Progress in Commonsense Reasoning The challenge was proposed explicitly as an alternative to the Turing test. However, modern language models have largely saturated it, which has shifted attention to harder benchmarks.
The Abstraction and Reasoning Corpus, known as ARC-AGI, takes a fundamentally different approach. Instead of language, it presents visual pattern-completion tasks that require what psychologists call fluid intelligence: the ability to reason about novel problems using minimal prior knowledge.13arXiv. Generalized Planning for the Abstraction and Reasoning Corpus Each task is unique, so a system cannot rely on memorized solutions. ARC-AGI has become a key benchmark in the AI research community precisely because pure machine-learning methods, including large language models, struggle with it.14arXiv. The ARC of Progress towards AGI: A Living Survey of Abstraction and Reasoning An upgraded version, ARC-AGI-2, was introduced to provide even more granular measurement of abstract reasoning at higher difficulty levels.15arXiv. ARC-AGI-2: A New Challenge for Frontier AI Reasoning Systems
Another school of thought argues that any purely text-based or screen-based test is inherently limited. A proposal for a “robotic Turing test” suggests that the real target for cognitive science should be explaining the full range of human performance capacity, including physical interaction with the world, not just the ability to hold a conversation.16PubMed. First, scale up to the robotic Turing test, then worry about feeling A system that can chat convincingly but cannot navigate a kitchen or hand someone a glass of water is, by this standard, missing most of what intelligence actually involves.
When Passing Does Not Mean Understanding
One of the more interesting findings from recent Turing test research involves what happens under the hood of the language models that pass. Studies examining pragmatic competence, the ability to understand implied meaning, read social context, and grasp what a speaker means beyond what they literally say, have found that large language models do show sensitivity to these cues, and that this sensitivity improves with scale.17ACL Anthology. The Pragmatic Mind of Machines: Tracing the Emergence of Pragmatic Competence in Large Language Models Reinforcement learning from human feedback, the process used to fine-tune chatbots after initial training, contributes further gains in scenarios that require reading between the lines.
This creates an uncomfortable ambiguity. On one hand, the models are clearly not “understanding” language in the way a human does; they have no lived experience, no body, no emotional life. On the other hand, they are doing something that produces outputs remarkably similar to what understanding would produce. The Turing test, by design, cannot distinguish between these two possibilities. It was built to be agnostic about internal states and to focus entirely on external behavior. That was elegant in 1950, when no machine could fake the behavior convincingly. Now that some can, the agnosticism starts to look like a blind spot.
Critics who argue the test primarily assesses “deceptive mimicry rather than genuine intelligence” have a point, but defenders counter that the test remains relevant precisely because it captures something important about how humans evaluate intelligence in practice.18arXiv. The Turing Test Is More Relevant Than Ever We do not, in everyday life, evaluate intelligence through abstract reasoning tasks. We evaluate it through conversation, through social interaction, through whether someone seems to “get it.” The Turing test mirrors that everyday process, for better and worse.
What the Persona Effect Reveals
Perhaps the most practically significant finding from the 2025 studies is the enormous gap between prompted and unprompted performance. GPT-4.5 without a persona prompt was judged human only 38% of the time. With a persona prompt, that jumped to 73%.19PubMed Central. Large language models pass a standard three-party Turing test The underlying model was identical in both conditions. The only difference was a set of instructions telling it to behave like a specific kind of person.
This has implications well beyond academic testing. In commercial applications, the same language model can be made to feel much more or much less human depending on how it is configured. A company that wants its customer-service chatbot to pass for human can do so with relatively simple prompt engineering. A company that wants its chatbot to be clearly identifiable as AI can achieve that too, by stripping away the persona cues. The technology is essentially neutral; what determines whether people are fooled is a set of design choices that sit on top of the model.
For regulators and ethicists, this raises a question that Turing never anticipated. The original framing assumed that passing the test would represent a fundamental technological achievement, a sign that machines had reached some meaningful threshold of capability. Instead, passing turns out to depend heavily on a thin layer of social packaging. The capability threshold that matters is not “can the machine think?” but “can the machine be configured to seem like a person?” And the answer to that second question, for current frontier models, is already yes.
How Typing Speed and Typos Tip the Scales
One underappreciated detail in Turing test experiments is the role of non-content signals. When conversations happen in real time, the speed at which responses appear becomes a cue. A human takes time to type, pauses to think, makes typos and corrects them. Early chatbots gave themselves away by responding instantly with perfectly formatted text. Modern test designs sometimes introduce artificial delays to neutralize this, but not all do, and in real-world interactions these timing cues remain a factor.
The Finnish study illustrates a related phenomenon: judges latched onto the use of colloquial language as proof of humanity.20arXiv. Cultural Competence in Context: A Large Language Model Passes the Turing Test in Finland Slang, regional expressions, and informal abbreviations served as human identity markers in ways that factual accuracy did not. This means a model trained heavily on formal text, academic papers, news articles, and textbooks, may perform worse in a Turing test than one trained on casual internet conversation, even if the former “knows” more in any meaningful sense. The test rewards fitting in socially, not knowing things.
For anyone designing or evaluating these tests, the lesson is that controlling for surface-level cues is genuinely difficult. Each time researchers eliminate one confound (typing speed, formatting, response length), judges find another proxy. The human ability to detect subtle social signals is powerful but also unreliable; it leads people to flag perfectly human-written responses as robotic because they were “too polite” and to accept AI-generated responses as human because they included a well-placed “haha.” The test ends up measuring a moving target, shaped as much by the judges’ assumptions as by the system’s abilities.