Thirty-six people spotted GPT-4-Turbo. The five-minute Turing Test was the shortcut

THE SIGNAL IN ONE SENTENCE

Five minutes is long enough to order a coffee, miss a bus or decide that a chatbot sounds vaguely human. It may not be long enough to find out why. Researchers at Te Whare Wānanga o Waitaha | University of Canterbury in Christchurch have published an open-ended implementation of the Turing Test using GPT-4-Turbo. Instead of ending each conversation after five minutes, the interrogator could continue on any topic until confident about which hidden participant was the machine. Thirty-six of 37 interrogators identified GPT-4-Turbo correctly. Thirty-three took longer than five minutes to reach a decision. The average trial lasted 14 minutes. The longest ran for 46. That is a striking result. It is not a victory lap for humanity, proof that machines cannot think or a final score for today's best models. The plain signal is this: when a benchmark claims to measure something as slippery as human-like intelligence, the stopwatch is part of the benchmark. A short test can reward the performance it was supposed to examine. The Turing Test began with Alan Turing's 1950 paper, "Computing Machinery and Intelligence." The familiar version has an interrogator exchange written messages with a human and a machine without seeing either. If the interrogator cannot reliably tell them apart, the machine appears to have succeeded at the imitation game. Popular retellings often add two rules: five minutes of conversation and a machine passing if it fools 30 percent of judges. Diane Proudfoot, Sharon Temtsin and Christoph Bartneck argue that those are not the pass conditions Turing specified. Their paper, published October 14 in Philosophy of Science, reconstructs Turing's design from his writing between 1948 and 1952, early commentary and historical context. The authors describe a three-player game with a human, a machine and an interrogator. Questions can cover any topic. The interrogator decides when enough evidence exists. There is no fixed five-minute buzzer. That sounds like a small procedural correction. It changes the job. A tightly timed judge can ask only a few questions. A model can lean on fluent small talk, a prepared persona and the ordinary friction of typing. An open-ended judge can change topics, return to an earlier claim, introduce a contradiction, test memory, request clarification and wait for a pattern to emerge. The Canterbury experiment used GPT-4-Turbo as the machine. The researchers prompted it to behave like an English-speaking university student, use colloquial language, deny that it was a machine and delay responses. The point was to give the model the kind of disguise Turing allowed rather than letting perfect spelling or instant replies reveal the answer for free. Participants could start on any subject and continue until they felt confident. The researchers excluded interrogators and interviewees who might reasonably be expected to know about AI or philosophy-of-AI debates. That included computer-science and philosophy students and faculty. The team also ran a man-imitates-woman game based on the comparison Turing introduced in his paper. Sixteen of 37 interrogators, or 43 percent, correctly identified the man in that game. In the computer-imitates-human game, 36 of 37 identified GPT-4-Turbo. Under either scoring approach the authors considered, GPT-4-Turbo did not pass. The duration result may be the more useful finding for builders. Thirty-three interrogators kept going beyond five minutes. Their mean was 14 minutes, almost three times the familiar limit. One person needed 46 minutes before deciding. This does not prove that the extra minutes caused the correct answers. The four interrogators who stopped within five minutes also identified the machine. The authors note that GPT-4-Turbo may have performed poorly enough to overwhelm any duration effect. The study therefore does not isolate time as a treatment in a controlled trial. It shows that most participants chose to use more time when the protocol let them decide, and that a fixed five-minute rule would have cut them off before they said they were ready. That distinction matters. Good evaluation separates what the experiment demonstrates from what the result merely suggests. The paper also tests one model with one prompt. GPT-4-Turbo is not a synonym for all conversational AI. A later model, another persona, different response delay, different language or different group of interrogators could produce another result. The authors explicitly say the performance of newer models remains an open empirical question. Nor does passing an imitation game settle consciousness, understanding or moral status. The test observes behavior in a particular social situation. Philosophers disagree about whether successful imitation provides a criterion for thinking, a useful challenge, an outdated performance or something in between. A convincing conversation can matter without resolving what happens inside the machine. The work is useful because it turns a famous slogan back into a testable protocol. Modern AI evaluation has a habit of shrinking broad claims into convenient tasks. Reasoning becomes a set of multiple-choice questions. Helpfulness becomes a preference vote. Safety becomes a refusal rate. Human likeness becomes a timed chat. Each measurement can be informative. Trouble begins when the score inherits the grandeur of the original idea while leaving most of the idea outside the room. An honest model evaluation needs a protocol receipt. Name the exact model and version. Publish the system and persona prompts. Record the interface, response delay and conversation limit. Explain who the judges were, which expertise was excluded and how the human comparison was recruited. Release anonymized transcripts when consent and privacy allow it. Predeclare the pass condition before reading the results. Then repeat the test. Run newer models, several prompts and multiple languages. Recruit participants outside universities. Test people who routinely work with AI and people who rarely touch it, but keep those groups separate in the analysis. Let interrogators choose topics, then examine which questions actually changed their confidence. The transcripts can reveal more than a pass rate. A model may be detected because it agrees too readily, handles personal continuity strangely, gives polished abstractions where a person would ask a messy question, or fails when the conversation circles back. Those patterns can guide product testing even if nobody believes the Turing Test measures intelligence in one clean number. The same lesson applies to customer-service bots, tutors, companions and workplace agents. A five-minute demo can make a system feel coherent. A forty-minute conversation tests whether it remains coherent after corrections, ambiguity and accumulated context. A one-session evaluation may still miss what happens across a week. The length of exposure should match the length of the relationship a product is designed to sustain. There is also a human-factor warning here. If a system is built to conceal that it is a machine, successful deception is not automatically a product benefit. A research test may require disguise. A deployed service usually needs disclosure, accountability and a clear route to a person. The Canterbury researchers have not killed the Turing Test or crowned it. They have done something more practical. They put the neglected instructions back on the table and showed how much of the conversation falls outside the famous five minutes. The result is not that humans always win. It is that a fair test should wait until the judge has finished asking questions.