On several formal tests, AI systems already outscore most people: Stanford's 2026 AI Index reports that some models meet or exceed human baselines on PhD-level science questions, multimodal reasoning, and competition mathematics. The same report shows the same systems failing simple tasks, such as reading an analog clock, where the top model is right only about half the time. "Outsmart" therefore depends on the test: AI is ahead on some narrow, well-defined problems and behind on others, and no benchmark measures general intelligence.
Key takeaways
- The 2026 AI Index says several models now meet or exceed human baselines on PhD-level science questions, multimodal reasoning, and competition math.
- Performance is uneven: a system can earn an International Mathematical Olympiad gold medal and still misread a clock face.
- Time matters. On one research-engineering benchmark, AI beat human experts with a two-hour budget, while people won by two to one at 32 hours.
- Hard exams written to defeat AI, such as Humanity's Last Exam, showed low scores when first reported, and leaderboards keep moving.
- A benchmark score describes performance on that benchmark, so it cannot show that a machine is smarter than a person in general.
Where AI has caught up with people on tests
The 2026 AI Index Report from the Stanford Institute for Human-Centered AI, with data through March 2026, states that several models now meet or exceed human baselines on PhD-level science questions, multimodal reasoning, and competition mathematics. It also reports that performance on SWE-bench Verified, a coding benchmark built from real software issues, rose from about 60 percent to nearly 100 percent in one year. The report notes that an AI system earned a gold medal at the International Mathematical Olympiad.
These are real results, and they are narrow ones. Each test has a fixed format, a known scoring rule, and a human baseline that someone had to define. Passing a science exam is not the same job as running a laboratory, and fixing a listed software issue is not the same job as maintaining a product for years.
Where people still lead
The same 2026 report says the top model reads analog clocks correctly only about 50 percent of the time, and it describes a "jagged" frontier in which gold-medal math sits beside basic perception errors. On OSWorld, a benchmark of computer-use tasks, agents reached about 66 percent in the report's chart against a human baseline of about 72 percent, and agents still fail roughly one attempt in three on structured tasks.
The 2025 edition of the index gives a sharper picture of how the comparison changes with effort. On RE-Bench, a set of research-engineering tasks, top AI systems scored four times higher than human experts when both had a two-hour budget. As the budget grew, people overtook the machines, and at 32 hours humans outscored AI two to one. The same edition reported that AI systems reached 35.5 percent on BigCodeBench, a coding test where the human standard was 97 percent.
Exams built to be hard for AI
Benchmarks saturate quickly, so researchers write harder ones. Humanity's Last Exam, described in a paper first posted in January 2025, contains 2,500 expert-written questions with closed-ended answers across mathematics, the humanities, and the natural sciences. The authors reported low accuracy and poor calibration for leading models and a significant gap between those models and the expert human frontier on this kind of question. The 2025 AI Index put the top score at 8.80 percent, and the 2025 edition also listed 2 percent on FrontierMath problems.
Scores on such exams rise over time, and any figure printed here would age. The more durable point concerns what the exam covers: closed-ended academic questions with a single checkable answer. That is a useful measure of expert knowledge recall and reasoning. It does not test long projects, judgment under uncertainty, or work with other people.
Summary of the tests discussed
| Test | What it measures | Reported result | What it leaves out |
|---|---|---|---|
| PhD-level science, multimodal reasoning, competition math | Answering structured questions | Several models at or above human baselines (AI Index 2026) | Open-ended research and teamwork |
| Analog clock reading | Basic visual perception | About 50 percent for the top model (AI Index 2026) | Shows uneven skill, not overall level |
| RE-Bench | Research-engineering tasks under a time budget | AI ahead at two hours, people ahead two to one at 32 hours (AI Index 2025) | Tasks far longer than a workweek |
| Humanity's Last Exam | Closed-ended expert questions | Top score 8.80 percent in the 2025 index | Anything without a single right answer |
Questions to ask about any "AI beats humans" headline
- Which humans formed the baseline: random volunteers, students, or experts in that field?
- How much time did each side get?
- Was the test public before the model was trained, so that answers could have leaked into training data?
- Does the task resemble real work, or only a slice of it?
- What happens on the nearest similar task the benchmark did not include?
Whether any of this adds up to a machine that outthinks people in general is a separate question on which researchers disagree, and the measurements above do not settle it either way.
FAQ
Has AI already outsmarted humans?
On specific tests, yes. The 2026 AI Index reports models at or above human baselines on several demanding exams. On other tasks, including basic perception and long multi-step work, people still lead, so the answer depends on the task.
Does winning a math olympiad mean a system is generally smarter?
No. A gold medal shows strong performance on olympiad problems. The same index records a top model misreading analog clocks about half the time, which illustrates how uneven the skills are.
Why do AI scores on hard exams change so fast?
Models improve, developers tune them against public benchmarks, and test questions can circulate online. A score is a dated measurement for one model version on one test.
Is there a test that measures general intelligence?
No agreed one exists. Researchers have proposed definitions and benchmarks, but each captures part of the picture, and the sources reviewed describe them as proposals.
Related reading
- Whether AI use makes people vulnerable or lazy
- Can AI experience emotions
- Where AI emotion recognition goes wrong
- Creative thinking tasks where top humans lead
More in AI Policy and Ethics.
Last verified: October 2026
Sources
- The 2026 AI Index Report, Stanford Institute for Human-Centered AI, 2026
- The 2025 AI Index Report, Technical Performance chapter, Stanford Institute for Human-Centered AI, 2025
- Humanity's Last Exam, Phan L et al., arXiv:2501.14249, 2025 (Nature, January 2026)