The Turing test is a behavioral test of machine intelligence in which a human interrogator holds text conversations with a hidden machine and a hidden person and tries to decide which is which. If the judge cannot reliably tell them apart, the machine is said to pass. Alan Turing proposed it in 1950 under the name imitation game, and it remains one of the best-known ideas in the AI fundamentals of artificial intelligence.
Turing deliberately replaced the vague question “Can machines think?” with an operational criterion based on observable conversation. That move made the test influential, but it also made it controversial, because passing depends on imitating human behavior, not on demonstrating understanding, reasoning or general capability.
How the imitation game works
The Turing test works by restricting all interaction to text, so that voice, appearance and handwriting cannot reveal which participant is the machine. In the original setup, an interrogator questions two hidden witnesses through a teletype channel, and the machine succeeds if it persuades the judge often enough that it is the human.
Turing did not define a precise pass mark. He predicted that by around 2000 an average interrogator would have no more than a 70 percent chance of a correct identification after five minutes of questioning. Later organizers turned this prediction into thresholds, which is why reported “passes” vary so much in meaning.
The table below lists the components of the classic protocol and what each one controls.
| Component | Role in the test | Typical choice |
|---|---|---|
| Interrogator | Asks questions and gives the verdict | Lay person or expert judge |
| Machine witness | Tries to be judged human | Chatbot or language model |
| Human foil | Provides the real human baseline | Volunteer answering honestly |
| Channel | Hides nonverbal cues | Text chat only |
| Time limit | Bounds the depth of probing | Five to twenty-five minutes |
| Pass criterion | Defines success | Misidentification rate near chance |
Variants of the Turing test
Several variants of the Turing test exist because the original description leaves room for interpretation. They differ in the number of participants, the expertise of the judges and the modalities that the machine must handle.
| Variant | Setup | What it adds |
|---|---|---|
| Three-party test | Judge compares a machine and a human | Direct comparison with a human baseline |
| Two-party test | Judge talks to one hidden partner | Simpler, but no paired comparison |
| Total Turing test | Adds vision and physical manipulation | Requires perception and robotics |
| Reverse Turing test | Machine judges whether a user is human | Basis of CAPTCHA systems |
| Expert-judge test | Judges trained to probe weaknesses | Harder, more informative verdicts |
The reverse Turing test is the most widely deployed form today. CAPTCHA challenges invert the roles, so that software tries to detect humans by tasks that are easy for people and assumed to be hard for automated programs.
How the Turing test compares with modern AI evaluation
The Turing test measures indistinguishability from humans, whereas modern AI evaluation measures task performance against ground truth. Benchmarks for reasoning, coding or knowledge score answers objectively, while the Turing test depends on a subjective human verdict that can be swayed by style, humor or deliberate errors.
The following comparison shows why researchers rarely use the test as a scientific benchmark, even though it remains central to public debate about artificial general intelligence.
| Approach | What is measured | Main weakness |
|---|---|---|
| Turing test | Human-likeness in conversation | Rewards deception and imitation |
| Task benchmarks | Accuracy on defined problems | Saturation and data contamination |
| Human preference ratings | Perceived answer quality | Bias toward fluent, confident text |
| Winograd schema tests | Commonsense pronoun resolution | Solvable by statistical shortcuts |
| Capability evaluations | Specific dangerous or useful skills | Narrow scope per evaluation |
The test is also closely tied to John Searle’s Chinese room argument, which claims that a system could pass by manipulating symbols without understanding them. This distinction between behavior and understanding is the core of the philosophical dispute, and it links the test to the symbol grounding problem.
Notable attempts and claimed passes
Claims that a machine has passed the Turing test are frequent, but most rely on short sessions, lenient thresholds or persona tricks. ELIZA in the 1960s showed that simple pattern matching could convince some users, an effect now called the ELIZA effect.
- The Loebner Prize ran annual competitions from 1991 to 2019, and no entry won the gold award for full indistinguishability.
- In 2014, the chatbot Eugene Goostman was reported to pass by posing as a 13-year-old non-native speaker, which excused its errors and drew wide criticism.
- Controlled studies of large language models in the 2020s reported that some systems were judged human at rates near or above chance in short three-party games, though results depend heavily on prompting and judge experience.
These episodes show that passing is a property of the test design as much as of the machine. A short session with untrained judges is far easier than a long interrogation by experts who know which questions expose machines.
Limitations and common misconceptions
The main limitation of the Turing test is that it measures human gullibility as well as machine ability. A system can succeed by being evasive or by imitating typing errors, and a highly capable system could fail simply because it answers too quickly or knows too much.
| Misconception | Why it is wrong | More accurate view |
|---|---|---|
| Passing proves consciousness | The test observes only behavior | It says nothing about inner experience |
| There is one official test | Turing gave no exact protocol | Results depend on the chosen variant |
| Failing means low intelligence | Non-human strengths can reveal a machine | Human-likeness is not capability |
| A pass equals general intelligence | Conversation covers limited skills | AGI needs broad task competence |
Despite these weaknesses, the test remains useful as a thought experiment about operational definitions and as a practical reminder that convincing conversation can mislead users about what a system actually knows.
The idea in brief
The Turing test is a guessing game. A person chats by text with two unseen partners, one human and one computer, and tries to work out which is which. If the person guesses wrong about as often as right, the computer has done a convincing impression of a human.
The key point is that the test checks how human a machine seems, not how smart or how aware it really is. A good actor can play a doctor without knowing medicine, and a program can sound human without understanding what it says.
A practical scenario
A university research group wants to know whether a new customer service chatbot sounds natural. It recruits 60 volunteers as judges, and each judge holds two five-minute text chats: one with the chatbot and one with a trained human agent, in random order. Afterward, each judge names the partner they believe was human.
The chatbot is picked as human in 27 of 60 sessions, or 45 percent, which is close to the 50 percent expected from pure guessing. The team reports that the bot is hard to distinguish in short exchanges, but it also notes that judges who asked about recent local events or multi-step arithmetic identified the machine far more often. The study therefore guides further testing of factual reliability instead of being treated as proof of intelligence.