- AI benchmarks become useless when models score too high to differentiate
- Humanity's Last Exam uses 2,500 expert questions frontier models cannot answer
- AI models fail HLE confidently, revealing a dangerous calibration gap
In 2020, Dan Hendrycks proposed a test designed to humble the most powerful AI systems in the world. It was launched in 2021.
Key figure
2021
The year Dan Hendrycks released MATH
The test was called MATH, and it consisted of competition-level mathematics problems that doctoral students found genuinely difficult. At the time of release, the best AI models scored below 10%. Hendrycks, a machine learning researcher at the Center for AI Safety in San Francisco, thought it would hold as a meaningful benchmark for how LLMs solved technical math issues, for years.
It held for three.
Hendrycks and team also designed a broader benchmark, MMLU (Massive Multitask Language Understanding), a broad collection of multiple-choice questions spanning 57 academic subjects. It was demanding when it launched.
By late 2024, leading models were scoring around 90% on these tests routinely.
The failure of these test (or rather, AI's ability to learn how to solve them, and others like it), led Hendrycks to build Humanity's Last Exam, the hardest AI benchmark ever attempted.
"They're now crushed," Hendrycks told Reuters in 2024, after OpenAI's o1 model swept through the standard battery of reasoning benchmarks.
They're now crushed.
Dan Hendrycks, Center for AI Safety, on traditional AI benchmarks
The Instrument Problem
A major challenge in the AI field is how to build instruments that measure AI skills - when AI skills evolve faster than the instruments are able to follow.
In science, measurement is everything. Physicists don't argue about whether particles are real; they argue about how to build devices sensitive enough to detect them. The same logic governs AI research.
Benchmarks are AI's instruments for comparison - their digital tape measure, so to speak.
If you cannot measure capability precisely, you cannot know what you have built or what it can do safely. You cannot know how close it is to doing something dangerous.
A saturated benchmark is worse than useless. It tells you the models are good. But it cannot tell you how good, where they fail, or what risks might be emerging just beyond what you can see.
What is benchmark saturation?
A benchmark saturates when AI models score so highly that the test can no longer distinguish between them. MMLU went from a frontier challenge in 2021 to near-meaningless by 2024, with models scoring above 90%. A saturated benchmark still produces scores. They just stop telling you anything useful.
Building Humanity's Last Exam
Hendrycks set out to design something that would not saturate quickly. The result, built jointly with Scale AI and published in Nature in January 2026, is Humanity's Last Exam, or HLE.
According to a New York Times interview (paywalled, sorry -Ed.) with Hendrycks in January 2025, the name was supposed to be "The Last Stand", but it was considered too dramatic.
The name is still partly a provocation, partly a serious claim. The researchers believe this is the last closed-ended academic benchmark AI will need before the question of "expert-level knowledge" is effectively answered.
The design choices are worth understanding, because each one reflects a lesson from what went wrong before.
Old benchmarks failed partly because their questions were findable. A model trained on vast swaths of the internet could encounter a test question, or something very like it, in its training data.
HLE addresses this with a strict rule: every question must stump frontier models before it can be accepted. Researchers logged over 70,000 question attempts. Of the 13,000 that stumped the models, human expert reviewers with graduate-level credentials culled the list to 2,500 final questions.
The questions are also closed-ended, with unambiguous answers that can be verified automatically. Open-ended questions introduce subjectivity into grading, making it impossible to compare models cleanly over time. HLE deliberately trades breadth of question type for precision of measurement.
Key figure
1,000
subject matter experts contributed questions to Humanity's Last Exam.
Nearly 1,000 subject-matter experts from more than 500 institutions across 50 countries contributed questions. Some draw on specialized research experience not indexed anywhere online, problems that emerged from a contributor's own laboratory work or unpublished findings.
The point was to find territory where pattern-matching from training data offers no advantage.
What the Scores Actually Reveal
The accuracy numbers from the initial evaluation were striking. Most frontier models scored between 10% and 37%.
Google's Gemini 3 Pro Preview topped the HLE leaderboard at approximately 37% as of early 2026. The rest clustered well below 30%.
That gap confirms something unsettling. The apparent mastery on saturated benchmarks was, at least in part, an artifact of measurement failure rather than genuine capability.
But the more troubling finding concerns calibration. A well-calibrated system, when it does not know the answer, should tell you that it, in fact, does not know. A model scoring 10% on a test should express roughly 10% confidence in its answers, not 90%.
HLE measured this explicitly. Across all models, calibration errors exceeded 70%. The models were consistently wrong, and consistently confident about being wrong.
The models simply could not estimate their own ignorance.
Key figure
70%+
Calibration error across all frontier models. They get it wrong and express high confidence doing so.
OpenAI researchers have shown that this kind of overconfident error is mathematically inevitable under current training methods. It is not merely an engineering problem to fix.
This matters beyond the research community. Deploying AI in high-stakes settings, in medicine, law, scientific research, depends on knowing not just how often a model is right. It also depends on how well the model knows when it is wrong.
The Benchmark That Knows It Will Break
The HLE team is candid about what their instrument cannot do. It tests structured academic problems, not open-ended research or creative problem-solving. High scores on HLE would not indicate artificial general intelligence or autonomous scientific capability.
A July 2025 independent review by FutureHouse, an AI research nonprofit, found that roughly 30% of answers in the biology and chemistry sections may be incorrect. The HLE team partially accepted the finding, noting that questions at this level of specialization carry inherent expert disagreement.
They are also candid about the instrument's likely lifespan. Recent history shows that benchmarks move from impossible to saturated faster than anticipated.
More On AI Measurements
AI Consciousness Checklist: 5 Things Scientists Can Measure
AI consciousness now has a scientific checklist. 20 researchers attempt to turn neuroscience theories into testable indicators.
→The team designed HLE with a private held-out question set, kept separate from the public benchmark. A model cannot be trained to beat questions it has never seen.
Hendrycks has said that once models break 50% on HLE, it will be reasonable to conclude AI has reached expert-level academic performance.
That threshold has not been crossed. But when it is, the field will need a new instrument.
The researchers intend for HLE to hasten the transition to a different class of benchmarks. These would measure not what a model knows, but what it can do with knowledge in open, unpredictable environments. Those benchmarks do not yet exist in rigorous form.
HLE holds, for the moment. The calibration data tells us more than the accuracy scores do.
The gap between AI confidence and AI capability is itself a frontier worth measuring carefully.
Sources
- Primary Source: A benchmark of expert-level academic questions to assess AI capabilities (Nature, 2026)
- Additional Context:
- Humanity's Last Exam project site (lastexam.ai)
- HLE Leaderboard (Scale Labs)
- About 30% of Humanity's Last Exam chemistry/biology answers are likely wrong (FutureHouse)
- Humanity's Last Exam: The Test That's Humbling the World's Smartest AI (Let's Data Science)
Fact Check: Claim-by-Claim Verification Verified
All 14 major factual claims verified across two rounds of independent checking. Core claims about Humanity's Last Exam — its design, scores, calibration findings, and the FutureHouse review — are well supported by the Nature paper, the HLE leaderboard, and published independent analyses.
Commentary
- The calibration error claim ("exceeded 70%") understates the paper's finding of >80% RMS calibration error. This is conservative rather than misleading.
- HLE leaderboard scores are dynamic — Gemini 3.1 Pro variants may now score higher (~44%) than the article's ~37% figure for early 2026.
- MATH benchmark problems are described as "competition-level" in both the article and the paper, though some sources note these are high-school competition level specifically.
Sources used for verification
Academic/Peer-reviewed:
- A benchmark of expert-level academic questions to assess AI capabilities - nature.com
- Measuring Mathematical Problem Solving With the MATH Dataset - arxiv.org
- Measuring Massive Multitask Language Understanding (MMLU) - arxiv.org
- Why Language Models Hallucinate - openai.com
Other reliable sources:
- Humanity's Last Exam project site - lastexam.ai
- HLE Leaderboard - labs.scale.com
- FutureHouse HLE Review - futurehouse.org
- MATH-500 Leaderboard - artificialanalysis.ai
Fact-checked by Perplexity Sonar Pro on 2026-03-24
