Is AI Actually Intelligent? What IQ Tests Can and Cannot Measure

Artificial intelligence keeps posting eye-catching scores on tests designed for human minds, and the founder of the UN-backed AI for Good platform is the latest prominent voice urging the public to take the technology's upside seriously. But there is a question buried under every headline about machine smarts: what does "intelligent" actually mean, and can the tools we built to measure human intelligence tell us anything about a machine? The short answer is mostly no — and understanding why says a lot about how your own mind works.

What happened: the "AI is intelligent" debate went mainstream

In a recent interview with Geneva Solutions, the founder of AI for Good — the United Nations ITU platform that convenes an annual summit on beneficial uses of the technology — argued that alongside warnings about risk, "we also need to talk about artificial intelligence's positive potential." The comment lands in the middle of a public argument that has been running hot since large language models went mainstream: one camp describes these systems as genuinely intelligent, another describes them as sophisticated pattern-matchers, and both sides reach for test scores to make their case.

That instinct is understandable. Humans have spent more than a century building instruments to quantify intelligence, and it feels natural to point those instruments at machines. But the comparison is far shakier than the headlines suggest, and the reasons why are a useful lesson in what intelligence tests do — and do not — measure in people.

What IQ tests actually measure in humans

A modern IQ test is a standardized instrument normed on a human population: the average score is set to 100, and roughly two-thirds of people land within 15 points of that average. Under the hood, most tests sample several distinct abilities:

• Fluid reasoning — solving novel problems you have never seen before, such as completing an unfamiliar visual pattern.
• Crystallized knowledge — vocabulary, general information, and learned facts.
• Working memory — holding and manipulating information in your head.
• Processing speed — how quickly you handle simple cognitive operations.

Two design assumptions make these tests meaningful for people. First, the test-taker has never seen the specific questions before, so performance reflects ability rather than memorization. Second, the score is a comparison against other humans who took the same test under the same conditions. Both assumptions collapse when the test-taker is a machine.

Why machines break the yardstick

The contamination problem

Large language models are trained on enormous swaths of the internet — which includes IQ test items, puzzle books, test-prep forums, and discussions of the answers. When a model aces a question that resembles its training data, that is crystallized knowledge at planetary scale, not fluid reasoning. It is the equivalent of handing a student the answer key years before the exam. This is why serious AI researchers treat human-test scores for machines as marketing rather than measurement.

Moravec's paradox: the skills profile is inverted

Roboticist Hans Moravec observed in the 1980s that computers find hard things easy and easy things hard. Machines surpassed the best humans at chess when IBM's Deep Blue defeated world champion Garry Kasparov in 1997, and at Go when DeepMind's AlphaGo beat Lee Sedol in 2016. Yet tasks any healthy five-year-old performs effortlessly — recognizing a sarcastic tone, folding laundry, noticing that a joke landed badly — remain difficult for machines. Human intelligence is a package deal shaped by evolution; machine capability is a jagged profile of superhuman peaks and toddler-level valleys. A single number like an IQ score cannot describe a jagged profile.

Narrow excellence is not general intelligence

Every system that beat humans at a game did so at exactly one game. Deep Blue could not play checkers. What psychologists call general intelligence — the ability that makes a person who is good at one mental task likely to be decent at many others — is precisely the thing machine benchmarks keep failing to demonstrate.

The benchmark built to catch pattern-matching

In 2019, AI researcher François Chollet published the Abstraction and Reasoning Corpus (ARC), a test designed to resist the contamination problem. Each puzzle shows a few before-and-after grid examples and asks the solver to infer the underlying rule and apply it to a new case — novel tasks, closer in spirit to fluid-reasoning items on a human IQ test. Ordinary people solve most ARC puzzles; for years, AI systems barely scored.

That changed in December 2024, when OpenAI's o3 reasoning model scored 75.7 percent on the ARC-AGI benchmark, and 87.5 percent in a high-compute configuration, according to the ARC Prize Foundation. It was a genuine leap — and the fine print still matters. The high-compute run reportedly burned thousands of dollars of computation per task to solve puzzles a human handles in seconds for the cost of a sip of coffee. Chollet himself drew the distinction that matters: skill is applying known rules under known conditions, while intelligence is generating new skills under unknown conditions. The foundation followed up in 2025 with a harder successor, ARC-AGI-2, on which frontier systems initially scored in the single digits while ordinary humans continued to solve most tasks.

The scoreboard keeps moving, but the pattern holds: machines close each gap only after enormous, targeted effort — and each new test of genuinely novel reasoning reopens the gap.

Why this matters for your own thinking

The AI-intelligence debate is not just academic. Three practical takeaways follow from it.

First, treat test scores — human or machine — as measurements of something specific, not verdicts on worth. An IQ score summarizes performance on a defined battery of tasks relative to other people. It predicts some real outcomes, but it does not capture judgment, curiosity, honesty, or the practical reasoning that daily life actually demands. If a single number cannot fully describe a machine, it certainly cannot fully describe you. That gap between abstract test performance and applied, everyday reasoning is exactly why practical assessments exist — you can take the Real World IQ test to see how your reasoning holds up on scenarios drawn from real life rather than abstract puzzles alone.

Second, the skills AI struggles with are the ones worth sharpening. Novel problem-solving, skeptical evaluation of sources, and transferring a lesson from one domain to another are the most durable human advantages. They are also trainable. Decades of puzzle-construction work — reaching 180 million total solvers across more than 80 countries over 30 years, a body of work recognized by Guinness World Records — rests on one consistent observation: people get measurably better at pattern recognition and lateral thinking with regular practice.

Third, critical thinking about AI is itself an intelligence test. When a headline claims a chatbot has a 150 IQ, the intelligent response is the fluid-reasoning move: ask what was measured, how, and compared to whom. The contamination problem alone should make you discount any machine score on a test written for humans.

What happens next

Expect the benchmark arms race to continue: new tests designed to isolate genuine reasoning, followed by systems engineered to beat them, followed by debate about what the victory proved. Expect voices like the AI for Good founder's to keep pushing attention toward applications — medical imaging, accessibility, disaster response — where the "is it really intelligent?" question matters less than whether the tool works. And expect the definition of intelligence itself to stay contested, because it always has been, even when humans were the only test-takers. The most useful move for any individual has not changed: understand what your own mind does well, measure it honestly, and train it. If you want a starting point, Real World IQ's assessment focuses on the applied reasoning that no benchmark controversy can take away from you.

FAQ

Does AI have an IQ?

No, not in any meaningful sense. IQ scores are normed against human populations taking unseen questions under controlled conditions. AI models have typically been exposed to test-like material during training, and their abilities are jagged — superhuman in some narrow areas, weak in others — so a single human-style score does not describe them.

Is AI smarter than humans?

At specific, well-defined tasks — chess, Go, rapid text summarization — machine systems exceed the best humans. At general intelligence, which means handling genuinely novel problems across domains, humans remain well ahead. Benchmarks built to test novel reasoning, like ARC-AGI-2, still show ordinary people outperforming frontier AI systems.

What is the ARC-AGI benchmark?

The Abstraction and Reasoning Corpus, created by François Chollet in 2019, is a set of visual puzzles that require inferring a rule from a few examples and applying it to a new case. Because the tasks are novel, it is much harder to pass through memorization, making it a better probe of fluid reasoning than human IQ tests repurposed for machines.

What human skills does AI struggle to replicate?

Novel problem-solving under unfamiliar conditions, common-sense physical and social reasoning, transferring lessons across domains, and judgment calls involving context and values. These map closely to fluid intelligence and practical reasoning — skills that improve with deliberate practice.

Can you improve your own reasoning ability?

You can improve measurably at pattern recognition, logical deduction, and applied problem-solving through regular, varied practice — puzzles, structured reasoning exercises, and honestly assessing where you stand today. Crystallized knowledge also keeps growing throughout life.