AI systems already exceed human performance on some tightly defined tests, but not across the board. The 2026 Stanford AI Index reports that frontier models meet or exceed human baselines on PhD-level science questions, competition mathematics, and multimodal reasoning. The same report finds that the best model read analog clocks correctly about half the time, against about 90 percent for people, so surpassing humans is a matter of which task is measured.
Key takeaways
- "Exceeding human capabilities" is measured task by task, against a human baseline, and the baseline choice affects the result.
- At the 2025 International Mathematical Olympiad a Gemini model scored 35 points, a gold-medal result, up from a 28-point silver-level result in 2024.
- Computer-use agents reached 66.3 percent on OSWorld, still below a human baseline of 72.35 percent.
- The AI Index calls the uneven pattern "jagged intelligence" and says benchmarks are saturating within months.
- Benchmarks themselves are imperfect: the report found invalid-question rates between 2 percent and 42 percent in several tests.
What "exceed" can and cannot mean
Computers have beaten people at specific tasks for decades, so the useful question is how broad the overlap has become. A benchmark gives one number for one task, and a human baseline is usually an average or an expert score from a study. A model that beats that number has beaten it on that test, under that scoring, with that time limit. It has not necessarily matched a person who must handle a changing, open-ended job.
The Stanford Institute for Human-Centered AI publishes the AI Index each year to track this with data. Its 2026 technical performance chapter, covering results through March 2026, opens with the finding that capability is outpacing the benchmarks designed to measure it. That is a statement about measurement as much as about machines.
Where the AI Index shows models ahead, close, or behind
The table collects figures from the chapter. Where the chapter text and its charts differ slightly, the text value is used.
| Task | AI result | Human reference | Reading |
|---|---|---|---|
| International Mathematical Olympiad, 2025 | Gemini Deep Think, 35 points, gold, in natural language within the 4.5-hour limit | Gold-medal score under the competition's rules | Gold-level on this exam |
| OSWorld (computer-use tasks) | 66.3 percent, up from roughly 12 percent | 72.35 percent baseline in the chart caption | Close, still behind |
| ClockBench (reading analog clocks) | Top model about 50 percent | About 90 percent | Far behind on a simple task |
| Household tasks by robots | 12 percent success in real homes | No baseline given | Strong in simulation (89.4 percent on RLBench), weak in practice |
The chapter's own text and chart differ by a fraction of a point on the clock and agent figures, which is why the table rounds the clock result. The AI Index lists a model's gold-medal math score and its clock failure side by side as its seventh takeaway.
Why speed of progress is a separate question
The chapter reports that frontier models gained 30 percentage points on Humanity's Last Exam in a single year, and that agents moved from answering questions to completing tasks while still failing roughly one in three attempts on structured benchmarks. Fast gains on a benchmark do not show that the next task, or a broader skill, will fall as fast. They show only that the benchmark is closer to saturation. Forecasts of when machines will outperform people at everything are opinion surveys, not measurements, and are treated in a separate article on this site.
Reasons to treat benchmark claims carefully
The AI Index also reports that benchmark reliability is a concern. Rates of invalid questions ranged from 2 percent on MMLU Math to 42 percent on GSM8K, so a model can lose points on questions that are themselves wrong. The report also finds the top models converging closely: the four leading developers sat within 25 Arena Elo points of each other in March 2026. A gap that small makes single-number rankings a weak guide to what a model can do for a given job.
Questions to ask about a claim that AI beats humans
- Which task and which benchmark, and when was it run?
- Who forms the human baseline: the general public, trained workers, or top experts?
- Was the test run under the same time and tool limits for people and the model?
- Has the benchmark been checked for invalid questions or leaked answers?
- Does the result hold outside the test, for example in the field, as the household robot figure suggests it may not?
FAQ
Is AI smarter than humans now?
On several narrow tests it matches or beats a defined human baseline. On others, such as reading analog clocks or household tasks, it falls well short. Intelligence as a single scale is not what these measurements capture.
What does "jagged intelligence" mean?
It is the AI Index's label for uneven ability: a model can win a gold medal in competition math yet fail a task most people find easy.
Can AI improve itself beyond human control?
Systems already help design other systems within tasks that people define and score. No source in this article shows unsupervised self-improvement, and a related article covers that evidence in detail.
Which benchmark should be trusted?
None alone. The AI Index itself documents invalid questions and fast saturation, so results are best read across several tests and with the method in view.
Related reading
- Can AI evolve
- Will AI surpass human intelligence
- Artificial and natural intelligence compared
- Whether AI could be conscious
More in AI Policy and Ethics.
Last verified: October 2026
Sources
- Technical performance, AI Index Report 2026, Stanford Institute for Human-Centered Artificial Intelligence, 2026
- AI Index Report 2026, Stanford Institute for Human-Centered Artificial Intelligence, 2026