“The Turing test is a method proposed by mathematician Alan Turing in 1950 to evaluate whether a machine can exhibit human-like intelligence. In this text-based conversation, a human judge communicates with two unseen participants. If the judge cannot reliably tell the machine apart from the human, the machine passes the test.” – Turing Test – Artificial intelligence
The central issue is whether convincing linguistic behaviour provides sufficient evidence for intelligence. A machine can produce answers that resemble human conversation without possessing consciousness, intentions, grounded experience, or dependable reasoning. That distinction makes the test valuable as a philosophical probe, but insufficient as a complete engineering standard for modern artificial intelligence.
In substance, the procedure is a controlled comparison between a human participant and a machine. A judge communicates with both through text while their identities remain concealed, then decides which participant is human. The machine succeeds when the judge cannot identify it reliably above chance or under the rules selected by the organisers. Turing presented this arrangement in his 1950 paper Computing Machinery and Intelligence, replacing the vague question of whether machines can think with a behavioural question that could be discussed more precisely 1.
The design rests on an important methodological choice. It excludes appearance, voice, physical movement and direct access to the system, concentrating on what can be observed through replies. This reflects a broader principle in the philosophy of mind: other people are ordinarily judged intelligent through their behaviour rather than through direct inspection of private consciousness. Turing therefore argued for a form of fair treatment, asking that machines be assessed by comparable external performance rather than rejected merely because their internal construction is different 4.
What passing actually measures
Passing the test demonstrates that a system can manage a social interaction well enough to defeat a particular judge under particular conditions. It may require grammatical fluency, contextual memory, humour, strategic ambiguity, factual knowledge and the ability to respond to unexpected prompts. Those capabilities are practically relevant because many human-computer interactions depend on communication rather than physical action.
Yet the result is not a direct measurement of a single internal property called intelligence. The judge observes outputs, not the process that produced them. A system might generate plausible language by exploiting regularities in training data, using evasive answers, adopting a recognisable persona or exploiting the judge’s assumptions. Conversely, a highly capable system might fail because it answers too formally, refuses ordinary requests, lacks culturally familiar mannerisms or is tested in a narrow conversation that does not reveal its strengths.
This creates a problem of validity. The test measures human-likeness in dialogue, while intelligence may include perception, planning, scientific discovery, motor control, learning from limited evidence, causal understanding and reliable action. A system that performs brilliantly in text but cannot maintain a consistent model of the world would pass one dimension of evaluation while failing others. Recent discussion of large language models accordingly treats the test as a historical or supplementary measure rather than a definitive benchmark 5,6.
Why the protocol matters
There is no single universally binding version of the test. Results depend on the number of judges, conversation length, question selection, participant instructions, success threshold and whether judges know that machines are present. A short exchange can reward style and confidence; a long exchange can expose contradictions, memory failures or shallow explanations. The choice of human comparison also matters, because people vary in writing ability, specialist knowledge and willingness to attribute intelligence.
A frequently cited 2014 event involving the programme Eugene Goostman reported that the system persuaded judges it was human in more than a stated threshold of short conversations 10. That claim attracted attention, but it did not settle the philosophical question. A result from one event cannot establish general intelligence unless the protocol is transparent, independently replicated and robust to changes in judges, prompts and comparison participants.
Statistically, a fair interpretation requires more than a headline percentage. If a judge chooses between two participants, random identification would have an expected accuracy of 0,5. A study might define a machine as successful when the judge identifies it no better than a chosen threshold, but that threshold should be specified before testing. With n independent judgements and k correct identifications, the observed accuracy is k/n. Confidence intervals, judge effects and repeated conversations are needed to distinguish genuine performance from sampling noise.
Major objections
The best-known objection separates simulation from understanding. John Searle’s Chinese Room argument contends that following formal rules for manipulating symbols could produce appropriate answers without semantic comprehension. Applied to conversational machines, the objection says that successful dialogue may show only that inputs are transformed into convincing outputs. The debate remains unresolved because supporters of behavioural evaluation argue that intelligence is properly attributed through sustained functional performance, while critics insist that behaviour alone cannot establish meaning or conscious understanding 6.
A second objection concerns anthropocentrism. Human conversation is one expression of intelligence, not its universal definition. Some systems are designed for theorem proving, weather prediction, protein analysis or robotic control rather than imitation of a person. Requiring such systems to appear human may reward unnecessary deception and penalise useful specialisation. It also risks treating human quirks as evidence of intelligence, even when those quirks are merely stylistic.
A third objection concerns reliability. Human judges bring expectations and biases to the exchange, and their decisions may be influenced by spelling, politeness, topic familiarity or the system’s willingness to admit uncertainty. The Stanford Encyclopedia of Philosophy describes the test more carefully as a specific imitation game and notes that its purpose was to avoid disputes about the meaning of thinking, not to provide a complete scientific definition of mind 1.
Contemporary relevance
The test still matters because it identifies a genuine threshold in machine communication: whether a system can sustain interaction without being immediately recognisable as artificial. That question has practical consequences for customer service, education, companionship, fraud prevention and public information. If conversational systems can imitate people, users need clearer disclosure, provenance signals and evaluation methods that test factual reliability rather than surface fluency.
For present-day AI, the most defensible approach is plural evaluation. Dialogue tests can measure interactional naturalness, but they should be combined with tests of factual accuracy, calibration, long-context consistency, reasoning, robustness to adversarial prompts, privacy, bias, safety and performance on real tasks. A system should not receive a broad intelligence claim merely because judges mistake it for a person. Nor should failure in a human imitation game be treated as proof that the system lacks valuable cognitive abilities.
The enduring contribution is therefore conceptual rather than merely competitive. Turing shifted attention from unresolvable assumptions about machine inner life towards observable performance, making machine intelligence discussable without first settling consciousness. The same move has a limit: observable performance must be measured against the capability that matters. Human-like conversation remains informative, but it is only one window onto an artificial system’s competence, reliability and social effects 13.
References
1. The Turing Test – Stanford Encyclopedia of Philosophy – 2003-04-09 – https://plato.stanford.edu/entries/turing-test/
2. The Turing Test – Stanford Encyclopedia of Philosophy – 2003-04-09 – https://plato.stanford.edu/archives/sum2025/entries/turing-test/
3. [PDF] Testing the relevance of the Turing Test against modern LLMs. – http://ijream.org/papers/IJREAMV11I10129007.pdf
4. Alan Turing – Stanford Encyclopedia of Philosophy – 2002-06-03 – https://plato.stanford.edu/entries/turing/
5. The Turing Test and our shifting conceptions of intelligence – 2024-08-15 – https://www.science.org/doi/10.1126/science.adq9356
6. The Turing Test at 75: Its Legacy and Future Prospects – 2025-01-01 – https://www.computer.org/csdl/magazine/ex/2025/01/10897255/24uGRl1DvJC
7. Turing test – Wikipedia – 2001-08-26 – https://en.wikipedia.org/wiki/Turing_test
8. Post Turing: – https://arxiv.org/pdf/2311.02049v1
9. [PDF] UC Merced – eScholarship.org – https://escholarship.org/content/qt7k75p125/qt7k75p125.pdf
10. Turing Test success marks milestone in computing history – 2014-06-08 – https://archive.reading.ac.uk/news-events/2014/June/pr583836.html
11. Microsoft Word – AGI2020_Efimov_final_AE.docm – https://philarchive.org/archive/EFIPMBv1
12. Computing Machinery and Intelligence – Wikipedia – 2003-12-16 – https://en.wikipedia.org/wiki/Computing_Machinery_and_Intelligence
13. Overcoming Turing: Rethinking Evaluation in the Era of Large Language Models | Stanford Law School – https://law.stanford.edu/2023/11/16/overcoming-turing-rethinking-evaluation-in-the-era-of-large-language-models/
14. Alan Turing – Stanford Encyclopedia of Philosophy – 2002-06-03 – https://plato.stanford.edu/archives/spr2019/entries/turing/
15. Computing Machinery and Intelligence (1950) – R Discovery – 2004-09-09 – https://discovery.researcher.life/article/computing-machinery-and-intelligence-1950/ae18dbfc1bce32b18fe73bb2e7f20a43
