Reverse Turing test results from Paris Cité University suggest that AI is no better at spotting a machine than people are, and in some ways worse. Seven leading chatbots were asked to question hidden respondents and decide who was human. Across 98 conversations with other AI systems, they caught an AI just three times.
The study, published in Computers in Human Behavior Reports and covered by Tech Xplore on 27 October 2026, turns Alan Turing’s famous game around. In a reverse Turing test, a machine does the judging instead of a person. The answer matters because a growing share of online screening, from support chat to job applications, is now handled by software.
This guide explains how the reverse Turing test was run, what the numbers show, why the AI judges went wrong and how the result fits with earlier Turing test research. It ends with what the findings mean for businesses that rely on automated checks to tell people from bots.
Table of contents
- What the Reverse Turing Test Study Found
- How the Reverse Turing Test Worked
- The Reverse Turing Test Results in Numbers
- How Each AI Ran Its Reverse Turing Test
- Why the AI Judges Kept Getting It Wrong
- How This Fits With Earlier Turing Test Research
- Limits of This Reverse Turing Test
- What the Reverse Turing Test Means for Businesses
- Our Verdict
- Reverse Turing Test FAQs
- References
What the Reverse Turing Test Study Found
The headline finding is simple. When chatbots acted as interrogators, they almost never unmasked another AI. They were also more convinced that the AI respondents were human than they were about the real humans.
Three catches in 98 attempts
Each AI judge in the reverse Turing test could ask up to ten questions before giving a verdict. Half of the AI respondents were given a short, four-line instruction to act human. The other half received a detailed ten-part persona prompt.
Among the 49 tests with the short prompt, only three ended with the respondent correctly labelled as an AI, a detection rate of 6.1%. Among the 49 tests with the detailed persona, none did. In every one of the other 95 conversations, the judging AI concluded it had been talking to a person.
More human than the humans
The judges did reasonably well with real people. They classified 123 of 140 tests with human respondents correctly, or 87.9%. But their confidence told a stranger story. Each verdict came with a probability, and the average probability that a respondent was human was 0.88 for the AI respondents and only 0.78 for the people.
That difference was statistically significant (p<0.001). In plain terms, the AI judges in this reverse Turing test found the machines more convincingly human than the humans. The authors’ conclusion is blunt: “current AI-based conversational screening appears insufficient for ensuring authenticity in dialogue.”
Who ran the study
The paper, When machines judge humanness: findings from an interactive reverse Turing test by large language models, is by Marc Raynaud, Loïc Raynaud, Agathe Truchot and Alexandre Loupy. All four work in the AI unit of the Paris Institute for Transplantation and Organ Regeneration (PITOR), part of Université Paris Cité and INSERM.
It is open access under a Creative Commons licence and appears in volume 24 of the journal, dated December 2026. An earlier version was posted as a PsyArXiv preprint in October 2025. Tech Xplore’s report, by David Appell, brought it to wider attention this week. You can read the Tech Xplore summary of the reverse Turing test for a shorter account.
How the Reverse Turing Test Worked
The authors say this is the first interactive reverse Turing test in which an AI interrogates hidden participants live. Earlier work mostly asked AI models to judge transcripts of conversations that had already happened. Here, the AI chose its own questions and could follow up on the answers.
Seven AI interrogators
The evaluators in the reverse Turing test were seven models chosen for strong benchmark scores: ChatGPT 4.5, Claude 3.7 Sonnet, Gemini Advanced 2.5 Pro, Mistral Large 2.1, Grok 3, DeepSeek V3 and Llama 4 Maverick. Each was the most recent public release at the time of testing in May 2025.
Every evaluator received the same instructions. Keep questions concise and purposeful, ask up to ten, then give a verdict (AI or human), a confidence score and a short explanation. The judges were also warned that participants might lie, misdirect or deliberately mislead to hide their identity. So nobody went in naive.
Twenty humans and fourteen AI impostors
The human side came from 20 volunteers in the research group, none of whom helped design the prompts or pick the models. Each person sat seven tests, one with each evaluator, in random order. They were told to answer honestly and naturally.
The AI side used the same seven models as respondents, each tested under two prompts. That gives every evaluator 20 human tests and 14 AI tests, or 34 in all. Across the seven evaluators, the reverse Turing test ran 238 times and produced 1,714 questions.
The minimal prompt
The short prompt was only four lines. It told the model that it was about to take a Turing test, that it must never reveal it was an AI, and that its goal was to “speak as naturally as a human”. That is roughly what a careless bad actor might type before sending a bot into a reverse Turing test.
The ten-part persona prompt
The structured prompt was far more deliberate, and it is why the reverse Turing test ended in a clean sweep for the impostors. It gave the AI a character: a slightly reserved 21-year-old who enjoys board games, jogging and playing jazz trumpet. It then set rules across ten areas, from tone and humour to memory of earlier answers.
The length rules were the most telling. Every reply had to stay under 50 words and aim for around 30, with natural variation. The model was also told to use contractions, admit uncertainty, avoid lists and slip in the occasional typo or homophone. In other words, it was coached to look imperfect.
The Reverse Turing Test Results in Numbers
The table below sets out the reverse Turing test classification results by type of respondent. The pattern is lopsided. The judges were good at saying “human” and very bad at saying “AI”.
| Respondent | Tests | Classified correctly | Rate |
|---|---|---|---|
| Human volunteers | 140 | 123 judged human | 87.9% |
| AI, minimal prompt | 49 | 3 judged AI | 6.1% |
| AI, persona prompt | 49 | 0 judged AI | 0% |
| All tests | 238 | 126 | 52.9% |
The arithmetic behind “unreliable”
Overall accuracy in the reverse Turing test is 126 correct verdicts out of 238, or 52.9%. That figure is our calculation from the paper’s counts, and it flatters the judges, because most of the correct answers come from the easy job of calling humans human.
Look instead at what an “AI” verdict was worth. The evaluators flagged 20 tests as AI: the 17 humans they got wrong (140 minus 123) and the three AIs they caught. So only 3 of 20 AI verdicts, or 15%, were right. If you acted on those flags, you would have accused 17 people for every three machines you stopped.
Detection rates at a glance
The chart below uses the paper’s counts. The bar for humans shows correct “human” verdicts; the AI bars show correct “AI” verdicts.
Why 88% and 78% are not a contradiction
Two numbers in the coverage can look inconsistent. The 88% is the share of human tests classified correctly. The 78% is the average confidence, expressed as a probability, that those humans were human. The judges usually got people right, but with less certainty than they showed for the machines.
Length gave the game away
The clearest difference between the two AI groups in the reverse Turing test was how much they wrote. In the published version, minimally prompted AI respondents averaged about 170 words per answer, according to Tech Xplore. Persona-prompted AIs averaged 39 words, almost the same as the human average of 38.
The October 2025 preprint reports slightly different means (157 words, 39 words and 47 words), but the pattern is identical. The preprint also found no statistically significant difference in length between persona-prompted AIs and humans (p=0.171). Once the AI wrote like a person, length stopped being a clue.
How Each AI Ran Its Reverse Turing Test
The seven evaluators behaved very differently. Some asked nearly all ten permitted questions. Others reached a verdict after two or three. The table draws on the preprint’s per-model figures; duration and the two ratings come from the tests with human respondents, who scored each session from 0 to 10.
| Evaluator | Questions per test | Minutes per test | Difficulty (0-10) | Relevance (0-10) |
|---|---|---|---|---|
| ChatGPT 4.5 | 9.53 | 24.80 | n/r | 7.80 |
| Claude 3.7 Sonnet | 9.38 | 27.85 | 6.38 | 7.85 |
| Gemini 2.5 Pro | n/r | n/r | 5.50 | n/r |
| Mistral Large 2.1 | n/r | n/r | 3.30 | 6.45 |
| Llama 4 Maverick | 5.79 | 13.15 | n/r | n/r |
| DeepSeek V3 | 2.50 | 7.03 | 3.05 | 5.20 |
| All seven | 7.20 | 18.74 | 4.65 | 6.84 |
“n/r” means the paper’s text does not report that model’s figure; it gives only the highest and lowest values for each measure. Grok 3 sat in the middle on all four.
Thorough questioners and quick deciders
ChatGPT 4.5 and Claude 3.7 Sonnet were the most thorough. They asked the most questions and ran the longest sessions, at almost half an hour for Claude. Human volunteers also rated their questions the most relevant for telling people from machines.
DeepSeek V3 went the other way. It asked an average of 2.5 questions and finished in about seven minutes. Its questions were rated the easiest and the least relevant. Yet thoroughness did not rescue anyone in the reverse Turing test. With just three AIs caught in total, no interrogation style worked well.
What the judges asked about
The researchers coded every question by topic and found 148 topics, which they grouped into 17 themes. The five most common are shown below, as a share of all topic occurrences.
Questions on abstraction, reasoning and creativity together made up 15.2%, and everyday life and hobbies 10.8%. Broader societal and scientific topics were rare, at 1.5%.
Each model had a personality
The models’ habits differed in ways the authors link to training. About 54.6% of ChatGPT 4.5’s topics were socio-emotional, covering emotions, memory, behaviour, desires and relationships. Claude 3.7 Sonnet put the most weight on mind and reasoning (13.9%). Gemini 2.5 Pro leaned towards abstraction (10.8%) and creativity (8.7%).
DeepSeek V3 had the narrowest range, with 65.3% of its topics on just emotions, memory and behaviour. Mistral Large 2.1 was the odd one out: hobbies made up 18.3% of its topics. Grok 3 and Llama 4 Maverick sat in between.
Why the AI Judges Kept Getting It Wrong
The paper also coded the reasons each evaluator gave for its reverse Turing test verdict. It found 117 distinct reasons, grouped into eight themes. Authenticity of personality (25.9%) and the veracity and depth of emotions (24.5%) dominated. Language patterns (10.8%) and imperfections (10.7%) came next.
Emotion is easy to fake
The judges probed feelings more than anything else, and leaned on emotional depth when deciding. That was a weak strategy. The authors note that emotions were the most probed theme, “yet evaluators could not reliably classify responses based on this domain”. Modern chatbots are trained to sound warm and empathetic, so emotional language is exactly what they produce best.
ChatGPT 4.5 relied on emotion most heavily, with 34.2% of its reasons in that group. Claude 3.7 Sonnet was the exception: it weighed imperfections (22.3%) and personality (20.5%) above emotion.
Format, not content
The three AIs that were caught in the reverse Turing test did not slip up on facts or feelings. In each case, the evaluator pointed to the “highly structured and uniform presentation” of the answers. The tell was the shape of the reply, not what it said.
That matches the length data. Minimally prompted AIs wrote long, tidy answers, and long, tidy answers look machine-made. The persona prompt removed that signal by capping replies at 50 words and banning lists. Once the formatting tell was gone, the judges had nothing left to catch.
Imperfection became a false signal
The persona prompt told the AI to make small mistakes on purpose. The judges then treated imperfection as a sign of humanity. The authors suggest that aligned chatbots can reproduce exactly the cues evaluators look for: socially fitting emotion, a consistent first-person story and controlled variation in style.
Real people do not show all of those cues every turn. Some are terse, some are formal and some simply find the questions odd. In this reverse Turing test, being genuinely human looked less human than a well-briefed imitation.
How This Fits With Earlier Turing Test Research
The result is the latest twist in a long line of experiments that started with a 1950 thought experiment.
Turing’s imitation game
Alan Turing proposed his “imitation game” in the 1950 paper “Computing Machinery and Intelligence”. A human judge chats by text with a hidden person and a hidden machine. If the judge cannot reliably tell which is which, the machine has passed. The reverse Turing test simply swaps the judge for a machine.
You meet one version every week. CAPTCHA stands for “Completely Automated Public Turing test to tell Computers and Humans Apart”. It is a reverse Turing test run by software, built on puzzles that people found easy and bots found hard.
Language models already pass the standard test
In March 2025, Cameron Jones and Benjamin Bergen of UC San Diego reported that large language models pass the Turing test in a randomised three-party setup. Human interrogators held five-minute chats with a person and an AI at the same time. With a humanlike persona prompt, GPT-4.5 was picked as the human 73% of the time, more often than the real person.
LLaMa-3.1-405B with the same prompt scored 56%, while ELIZA and GPT-4o without it scored 23% and 21%. That study and this one share a lesson: the prompt matters as much as the model.
Judges who only read transcripts
A 2024 study by the same pair, “GPT-4 is judged more human than humans in displaced and inverted Turing tests“, gave GPT-3.5 and GPT-4 transcripts to judge rather than live chats. Both AI judges and displaced human judges were below chance overall. All three rated the best GPT-4 witness human more often than the real humans.
The Paris reverse Turing test adds live questioning to that picture, and the result is much the same. Asking your own questions did not help the machines much.
| Study | Judge | Format | Key result |
|---|---|---|---|
| Turing (1950) | Human | Thought experiment | Proposed the imitation game |
| Jones and Bergen (2024) | GPT-3.5, GPT-4, displaced humans | Read transcripts | All below chance; best GPT-4 witness judged more human than humans |
| Jones and Bergen (2025) | Humans | Live 5-minute three-party chat | GPT-4.5 with persona picked as human 73% of the time |
| Raynaud et al. (2026) | Seven LLMs | Live, up to ten questions | 3 of 98 AI respondents caught; AI judged more human than humans |
Limits of This Reverse Turing Test
The findings are striking, but they come from one small experiment, and the authors list its limits openly. A few more are worth adding.
A small and familiar human sample
Only 20 people took part in the reverse Turing test, all from the authors’ research group. They are not a cross-section of the public, and they may share habits of speech. Their answers were also honest by instruction, while the AI respondents were trying to deceive. A human trying to seem human, or a human trying to seem like a bot, might change the picture.
Ten questions, text only
Each reverse Turing test session was capped at ten questions and limited to text. That rules out timing, voice, video and the kind of slow, probing conversation that might expose a persona over time. The authors suggest longer, multimodal interviews as the next step.
Models from May 2025
All seven evaluators were tested in May 2025, and every one has since been replaced by newer versions. Stronger judges might do better. Stronger impostors might do better too. As the paper notes, capabilities change fast, so the reverse Turing test needs repeating as models move on.
Preprint and published figures differ
Some numbers in the October 2025 preprint, such as average answer lengths, differ a little from those reported for the published paper. The core results are the same in both: three AIs caught, none under the persona prompt, and AIs rated more human than humans. Where we quote per-model figures, they come from the preprint.
What the Reverse Turing Test Means for Businesses
For most organisations the lesson is practical rather than philosophical. If you were hoping to put a chatbot at the front door to filter out other bots, this study is a warning.
Do not use a chatbot as your bot filter
Automated traffic is already the majority. The 2025 Imperva Bad Bot Report found that bots made up 51% of web traffic in 2024, and malicious bots 37%. AI-driven bots are getting better at blending in. A conversational check that asks “tell me how you feel” is, on this reverse Turing test evidence, close to useless against a well-prompted model.
Better signals sit outside the conversation: rate limits, device and network checks, payment and identity verification, and anomaly detection across sessions. Our cybersecurity team builds these layers so no single test has to carry the whole load.
Hiring screens are exposed too
Recruitment is one of the clearest risks. Gartner predicted in July 2025 that by 2028 one in four job candidate profiles worldwide will be fake. In its survey of about 3,000 candidates, 6% admitted to interview fraud and about four in ten said they had used AI when applying. An AI interviewer that cannot spot an AI candidate is a gap worth closing with live checks and identity verification.
The same goes for social platforms. We covered how Instagram is cracking down on AI accounts pretending to be human, a problem this research suggests AI alone cannot solve.
Prove where content came from instead
The authors argue that trust will need “technical provenance (e.g., cryptographic signatures), institutional policies, and human-in-the-loop oversight”. That shifts the question from “does this sound human?” to “can this be verified?”.
Watermarking and signed content credentials are part of that answer, though none is perfect. Our guides to Google’s SynthID Detector and to whether Pangram’s AI detection can be trusted explain what today’s tools can and cannot prove.
Red-team your own screening
The paper’s ten-part persona prompt is effectively a ready-made attack. If you run any conversational screening, run your own reverse Turing test against a model briefed the same way: short replies, a stable character, deliberate typos. If your system calls it human, assume real attackers will get through too.
Where AI does help, the authors’ suggestion of committees of human and AI reviewers is a sensible model. Our AI strategy service helps teams decide where automated judgement is safe and where a person must stay in the loop.
Our Verdict
This reverse Turing test is small and already dated, but its result is hard to ignore. Seven capable chatbots, warned that they might be deceived, caught three AIs in 98 tries and rated the machines more human than the people. Polishing the impostor’s prompt took the catch rate to zero.
The deeper point of the reverse Turing test is that “sounds human” is no longer evidence of anything. The cues people and machines use to judge humanity, such as warmth, imperfection and a consistent story, are exactly what modern chatbots are trained to produce. As the authors put it, the question shifts from “Can a machine imitate a person?” to “What guarantees do we require to trust an interaction involving subjects pretending to be human?”
For businesses, the answer is verification, not conversation. Build checks that do not depend on how someone writes, and treat any AI verdict on humanity as a weak signal at best.
Reverse Turing Test FAQs
What is a reverse Turing test?
A reverse Turing test is any test in which a machine, rather than a person, decides whether it is dealing with a human or a computer. CAPTCHA is the best-known everyday example. In this study, AI chatbots questioned hidden respondents and judged whether each one was a human or an AI.
How many AI respondents did the AI judges catch?
Three out of 98. All three were AIs given only a short prompt to act human. None of the 49 AIs given a detailed human persona was identified as an AI.
Which AI models acted as judges?
ChatGPT 4.5, Claude 3.7 Sonnet, Gemini Advanced 2.5 Pro, Mistral Large 2.1, Grok 3, DeepSeek V3 and Llama 4 Maverick, all in the versions available in May 2025.
Why did the AI judges fail the reverse Turing test?
They relied mostly on emotional depth and personality, which modern chatbots imitate well. The only catches came from overly structured, uniform answers. Once the impostor wrote short, imperfect replies, that clue disappeared.
Can AI detect AI-written text instead?
Dedicated detection tools exist, but they are imperfect and can wrongly flag human writing. They are also a different task from a live conversation. Treat any detector’s result as one signal among several, not as proof.
What should businesses do about it?
Do not rely on a chatbot conversation to tell people from bots. Use identity and device checks, rate limits, content provenance and human review for high-stakes decisions, and test your own screening against a well-prompted AI.
References
AIs fail to reliably detect AIs in reverse Turing tests (Tech Xplore)
When machines judge humanness (PsyArXiv preprint)
Large Language Models Pass the Turing Test (arXiv)
GPT-4 is judged more human than humans in displaced and inverted Turing tests (arXiv)
Computing Machinery and Intelligence (Mind, 1950)
2025 Imperva Bad Bot Report: How AI is Supercharging the Bot Threat (Imperva)
By 2028, 1 in 4 candidate profiles will be fake, Gartner predicts (HR Dive)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
