IQ Metrics
IIF Certified Assessment Start IQ Test
IIF Certified Assessment
AI & Machine IntelligenceEmerging evidence

What Is an Artificial Intelligence Test? Three Meanings, Compared

An artificial intelligence test can be a Turing test, an AI benchmark or a skills test for people. Here is who is tested, what comes out and what, if anything, links each one to IQ.

What Is an Artificial Intelligence Test? Three Meanings, Compared
Illustration generated for IQ Metrics. No photograph is used. IQ Metrics

What you need to know

  • An artificial intelligence test can mean three things: a Turing test of whether a machine passes as human, an AI benchmark that scores a model on percent correct, or a skills test an employer gives a person. None of the three produces an IQ.
  • A Turing result is a win rate: in a 2026 PNAS study GPT-4.5 with a persona was picked as the human in 73% of five-minute chats and 36% without it. A benchmark is percent correct against an answer key, such as MMLU's 15,908 questions.
  • Only a human IQ test given to a model produces a count a person could compare. On the Mensa Norway test the mean for AI models rose from 12.0 right answers of 35 in May 2024 to 30.1 in September 2026, but that count is not an IQ.
  • The one place an AI score meets a human ranking is a human exam: at the 2025 International Mathematical Olympiad the only AI result the IMO confirmed placed joint 27th of 630 contestants.

An artificial intelligence test can mean three different things. It can be a Turing test, which asks whether a machine can pass as a human in conversation. It can be a benchmark, a fixed exam such as MMLU that scores an AI model on percent correct. Or it can be a skills test an employer gives a person. None of the three produces an IQ.

We did not find a page that sets the three side by side, so this one does, by who is tested, what comes out and what links to IQ. The short version: a Turing result is a win rate, a benchmark is a percent correct, and only a human IQ test gives an IQ.

Below: the three meanings, how each is scored, what AI models score today, and whether an AI can sit an IQ test. Every figure comes from a page we link or a data file we downloaded on 19 November 2026.

What is an artificial intelligence test?

The phrase covers things that share a name and little else. A test can be aimed at a machine or at a person. It can measure humanlikeness, accuracy or job skills. The table below puts them side by side.

  • Turing test: a chatbot is tested. The result is the share of conversations in which a judge picks it as the human. Link to IQ: none, because it measures humanlikeness
  • AI benchmark: an AI model is tested. The result is percent correct on a fixed set of questions. Link to IQ: none, because there is no human norm group and no standard deviation
  • AI skills test for hiring: a person is tested, for a job that uses AI. The result is a score on the employer's chosen test. Link to IQ: we read none of these tests, so we state none
  • A human IQ test given to a model: an AI model sits questions written for people. The result is a count of right answers. Link to IQ: only through a formula a tracker chose, not through human norms

What is the Turing test, and has AI passed it?

The Turing test asks whether a judge can tell a machine from a person in a text conversation. In a 2026 PNAS study, GPT-4.5 given a persona was picked as the human in 73% of five-minute chats. The same model without the persona managed 36%.

Alan Turing did not set a pass mark. In his 1950 paper he wrote that in about fifty years a computer could play the imitation game so well that an average judge would have no more than a 70% chance of identifying it correctly after five minutes of questions. That is a forecast, not a threshold.

The PNAS authors say the test measures humanlikeness, not intelligence. Our guide to whether AI has passed the Turing test covers the study, the persona prompt and the 15-minute follow-up.

How do you test an AI? What AI benchmark tests measure

How do you test an AI? You give the model a fixed set of questions with an answer key, and its score is the share it gets right. That is what AI benchmark tests do. MMLU has 15,908 multiple-choice questions in 57 subjects, so 25% is a random guess. Humanity's Last Exam has 2,500 closed-ended academic questions, kept only if frontier models failed them when it was built.

The ARC-AGI-3 benchmark is built differently. When it was announced on 25 March 2026, humans scored 100% and the best frontier AI scored 0.51%, because every task had first been solved by people. Our guides to MMLU, Humanity's Last Exam and ARC-AGI-3 give the details.

Even one benchmark can have several top scores. On 23 October 2026 three public boards for Humanity's Last Exam gave three different leading scores, 54.8%, 59.9% and 61.4%, because each used a different question set, judge or tool setting. They are different tests, so the figures should not be merged.

A benchmark score also depends on the rules. One lab reported the same model on one benchmark at 77.9% under partial scoring and 41.7% under strict scoring. Our piece on why AI benchmark scores disagree explains the gap. A benchmark has no age norms and no standard deviation, so its percent correct cannot be turned into an IQ.

A Turing result is a win rate, a benchmark is a percent correct, and only a human IQ test gives an IQ.

There is one test where an AI result meets a human ranking, because it is a human exam. At the 2025 International Mathematical Olympiad the only AI result the IMO confirmed, 35 of 42, would have placed joint 27th among 630 teenage contestants. Our AI versus the Olympiad study has the ranking and the problem no AI solved. A rank among contestants is still not an IQ.

Are there AI tests for people, and can an AI take an IQ test?

Yes to both. Employers give people skills tests when they hire for jobs that use AI, and an AI model can sit a human IQ test. The two results mean different things, and neither is an IQ for the machine.

An AI skills test for hiring checks what a candidate can do with AI tools. We did not read the technical documents of any of these tests, so we say nothing about their content, norms or scores, and we name no vendor. A different problem is AI sitting the test for the candidate. One platform reviewing 19,368 AI-led interviews flagged 38.5% for AI assistance, as our guide to AI agents and proctored hiring tests reports.

So can AI take an IQ test? It can sit the questions. TrackingAI's public log of the Mensa Norway test, which we downloaded on 19 November 2026, holds 1,931 AI sittings since 9 May 2024. The mean rose from 12.0 right answers of 35 in May 2024 to 25.2 in April 2026 and 30.1 in September 2026. Thirteen sittings, 0.67%, were perfect.

  • May 2024: 12.0 of 35 right on average, 94 sittings
  • May 2025: 19.3 of 35, 94 sittings
  • April 2026: 25.2 of 35, 84 sittings
  • September 2026: 30.1 of 35, 73 sittings

A count of right answers is not an IQ. The tracker rescales it with its own straight line, which turns 35 of 35 into 151, above the human test's top of 145. Mensa Norway publishes no such table. Our guides to the Mensa Norway IQ test and to AI scores on it explain why.

To put a human score in context, the IQ percentile calculator shows where an IQ lands among people. For the same kind of reasoning tasks on your own, the brain games include logic puzzles.

Your own number

Where would your own score land?

Want a score that compares you with people, not with a machine? Take the IQ Metrics test and read it against the bell curve.

Find your IQ score now! →
Secure & encryptedInstant results10–20 minutes

When you see an AI "test score" in a headline, run through five questions. They take a minute and settle most arguments.

  • Who or what was tested: a chatbot, a model, or a person?
  • What came out: a win rate, a percent correct, or a count of points?
  • Is there a human norm group, with a mean and a standard deviation? Without one there is no IQ
  • Which rules, tools and grading were used, and who ran the test?
  • When was the score read? Leaderboards change within weeks

For a result on the standard IQ scale, with a percentile you can check, take the IQ test.

Common questions

What is an artificial intelligence test?

The phrase has three common meanings: a Turing test of whether a machine passes as human, an AI benchmark that scores a model on percent correct, and a skills test employers give people for AI-related jobs. None of them gives an IQ.

How do you test an AI?

Mostly with a benchmark: a fixed set of questions and an answer key, scored as percent correct. MMLU has 15,908 questions in 57 subjects. Scores depend on grading rules, so the same model can score 77.9% under one rule and 41.7% under another.

Can AI take an IQ test?

It can sit the questions, but the result is a count of right answers, not an IQ, because no human norm group stands behind it. On the Mensa Norway test the mean for AI models rose from 12.0 to 30.1 of 35 between May 2024 and September 2026.

Has AI passed the Turing test?

In one 2026 PNAS study GPT-4.5 with a persona was picked as the human in 73% of five-minute chats, against 36% without the persona. The authors say the test measures humanlikeness, not intelligence.

Sources for this story

  1. Computing Machinery and Intelligence (A. M. Turing, 1950), Mind 59(236): the 70 per cent, five-minute forecast (text read on 19 November 2026) — Mind, Oxford University Press
  2. Large language models pass a standard three-party Turing test (Jones and Bergen), PNAS 123(21), 19 May 2026 — PNAS
  3. Measuring Massive Multitask Language Understanding (Hendrycks and colleagues), ICLR 2021: 15,908 questions, 57 subjects — arXiv / ICLR
  4. TrackingAI score log (daily_logs.csv), Mensa Norway sittings 9 May 2024 to 9 November 2026, downloaded 19 November 2026 — TrackingAI (trackingai.org)
  5. IQ Metrics analysis of the 2025 International Mathematical Olympiad results (joint 27th of 630) — IQ Metrics Research

Corrections: spotted an error? Email corrections@iqmetrics.org and we will update this story and note the change here.

Share this story

Know someone who keeps seeing this number quoted without the scale it was measured on? Send it to them — it takes one tap.

Filed under#ai benchmarks#fluid reasoning#study quality#work and hiring

Read the research.
Then find your own number.

Our IIF-certified assessment reports your score with its scale, percentile and confidence range — and a breakdown of the cognitive domains behind it.

Start IQ Test →
Secure & encryptedInstant results10–20 minutes