Can you beat AI?
At a glance
10 faculties · 2 estimated human refs · live sampleAI now aces exams that stump most people. But intelligence isn’t one number. Google DeepMind’s 2026 framework splits it into ten faculties, from perception to social cognition. Where does humanity stand? As of October 2026, on published benchmarks, people still lead AI on four of them, frontier models lead on three, and the other three are too close to call. For now.
Most “AI vs human” claims compare a model to a single number: the average person, or an expert. But people differ enormously. We measure the whole human distribution on every task, from the 1st to the 99th percentile and across age, education and background, so any AI system can be placed precisely: at the 34th percentile of adults on working memory, not just “below human”. Nobody has measured the whole human range on these tasks. We are.
This study measures where you sit. Seven short items, each modelled on a published AI benchmark, show your answer next to how frontier models did. Every answer joins an open dataset comparing human and machine cognition on the same tasks.
- Humans lead on
- 4/10
- Too close to call
- 3/10
- Largest human lead
- Attention +44.0
- Largest AI lead
- Metacognition -55.0
Cognitive profile
Best published AI vs human reference on each faculty, 0–100. Take the test to add your own line.
Both shapes are lopsided, and the bumps are in different places. AI pulls ahead on expert reasoning, producing work and flagging false premises.
People stay ahead on attention, memory, perception and executive control, the faculties that everyday life leans on.
Hollow points are our own estimates where no published human figure exists.
Human vs AI
Head-to-head by faculty, best published result on each side. Human figures are published results on other tests, or our estimates (~), until our own sample reaches 30 people per item.
| Faculty | Human | Best AI | Gap | Human reference | AI reference |
|---|---|---|---|---|---|
| Perception Taking in information from the world and making sense of it. | 91.4 | 78.4 | +13.0 | Perception Test, untrained adults | Perception Test, Gemini 2.5 Pro |
| Generation Producing words, images, sounds or actions, and executing them well, including under constraints. | 50.0 | 84.9 | -34.9 | GDPval, expert vs expert (parity by design) | GDPval wins or ties, GPT-5.5 (vendor-reported) |
| Attention Focusing on what matters and filtering out distractors that look relevant. | ~95.0 | 51.0 | +44.0 | Our estimate, untimed reading task | RIAC, one planted lure, pooled |
| Learning Picking up new rules or skills from examples or feedback, then applying them. | 100.0 | 99.9 | +0.1 | ARC-AGI-3, members of the public | ARC-AGI-3, GPT-6 Astra with custom harness |
| Memory Holding on to information over time and retrieving it when needed. | 87.9 | 66.5 | +21.4 | LoCoMo F1 (Maharana et al., 2024) | LoCoMo F1, Claude Sonnet + memory system |
| Reasoning Drawing valid conclusions: deduction, induction, analogy and maths. | 69.7 | 96.4 | -26.7 | GPQA Diamond, PhD experts | GPQA Diamond, Gemini 4 Argon (Epoch AI) |
| Metacognition Knowing what you know: spotting broken questions, judging your confidence. | ~45.0 | 100.0 | -55.0 | Our estimate (Moses illusion + invented names) | MetaBound, best model; two models at 0% |
| Executive functions Planning, switching strategy when rules change, inhibiting old habits. | 92.0 | 74.5 | +17.5 | GAIA, university-educated testers | GAIA, Claude Sonnet 4.5 (HAL) |
| Problem solving Combining faculties to solve problems, from novel puzzles to common sense. | 83.7 | 81.9 | +1.8 | SimpleBench, nine non-specialists | SimpleBench, Claude Fable 5 |
| Social cognition Understanding what people think, feel and mean without saying it. | 84.4 | 80.0 | +4.4 | Social IQa (Sap et al., 2019) | CogToM, GPT-5.1 (a different test) |
The standard benchmarks, passed one by one
Best AI result each year, as a share of each benchmark's human baseline (100%)
On the benchmarks AI research has leaned on since 2012, the best system has caught up with the human baseline one after another: 8 of 11 by 2025, and the newer ones get there in a year or two rather than five.
Each baseline is one number, often from a small group of people or experts, and it can differ from the figures we use. Once a benchmark is passed it says little about where AI stands against people, which is why this project tests the faculties where it doesn’t lead yet, and measures the whole range of people rather than one number.
Redrawn from the data of Stanford HAI, AI Index Report 2026, Fig. 2.1.1 (CC BY-ND 4.0), the latest edition; its last year is 2025. Hover or tap a panel to read each year’s best model.
Data table
| Task | Benchmark | Year | % of human | Best model |
|---|---|---|---|---|
| Image classification | ImageNet Top-5 | 2012 | 89.1 | AlexNet |
| Image classification | ImageNet Top-5 | 2013 | 91.4 | OverFeat - 7 accurate models |
| Image classification | ImageNet Top-5 | 2014 | 96.9 | VGG-19 |
| Image classification | ImageNet Top-5 | 2015 | 99.5 | Inception V3 |
| Image classification | ImageNet Top-5 | 2016 | 100.7 | ResNeXt-101 64x4 |
| Image classification | ImageNet Top-5 | 2017 | 101.4 | PNASNet-5 |
| Image classification | ImageNet Top-5 | 2018 | 102.8 | ResNeXt-101 32x48d |
| Image classification | ImageNet Top-5 | 2019 | 103.8 | BiT-L (ResNet) |
| Image classification | ImageNet Top-5 | 2020 | 104.1 | Meta Pseudo Labels (EfficientNet-L2) |
| Image classification | ImageNet Top-5 | 2021 | 104.3 | Florence-CoSwin-H |
| Image classification | ImageNet Top-5 | 2022 | 104 | Top-k DiffSortNets (EfficientNet-L2) |
| Visual reasoning | VQA | 2016 | 80.1 | MCB |
| Visual reasoning | VQA | 2017 | 86.5 | Image features from bottom-up attention (adaptive K, ensemble) |
| Visual reasoning | VQA | 2018 | 86.7 | BAN+Glove+Counter |
| Visual reasoning | VQA | 2019 | 90.7 | UNITER (Large) |
| Visual reasoning | VQA | 2020 | 91.4 | Oscar |
| Visual reasoning | VQA | 2021 | 102.5 | VLMo |
| Visual reasoning | VQA | 2022 | 104.4 | PaLI |
| Medium-level reading comprehension | SQuAD 2.0 | 2017 | 82.4 | SAN (ensemble model) |
| Medium-level reading comprehension | SQuAD 2.0 | 2018 | 96.2 | PAML+BERT (ensemble model) |
| Medium-level reading comprehension | SQuAD 2.0 | 2019 | 103.3 | ALBERT + DAAF + Verifier (ensemble) |
| Medium-level reading comprehension | SQuAD 2.0 | 2020 | 103.9 | SA-Net on Albert (ensemble) |
| Medium-level reading comprehension | SQuAD 2.0 | 2021 | 104.1 | IE-Net (ensemble) |
| English language understanding | SuperGLUE | 2019 | 94.2 | Roberta |
| English language understanding | SuperGLUE | 2020 | 99.4 | T5 Team - Google |
| English language understanding | SuperGLUE | 2021 | 101.6 | Liam Fedus |
| English language understanding | SuperGLUE | 2022 | 101.7 | Vega v2 |
| English language understanding | SuperGLUE | 2024 | 101.8 | Hairuo |
| Multitask language understanding | MMLU | 2019 | 36.1 | GPT-2 1.5B (fine-tuned) |
| Multitask language understanding | MMLU | 2020 | 60 | GPT-3 (fine-tuned) |
| Multitask language understanding | MMLU | 2021 | 66.8 | Gopher (few-shot, k=5) |
| Multitask language understanding | MMLU | 2022 | 83.7 | Flan-PaLM (5-shot, finetuned, CoT + SC) |
| Multitask language understanding | MMLU | 2023 | 96.2 | GPT-4 |
| Multitask language understanding | MMLU | 2024 | 102.8 | o1-preview |
| Competition-level mathematics | MATH | 2021 | 7.7 | GPT-2 (1.5B) |
| Competition-level mathematics | MATH | 2022 | 57.6 | GPT-4 model (w/ code, PAL) |
| Competition-level mathematics | MATH | 2023 | 93.7 | GPT-4-code model (CSV, w/ code, SC, k=16) |
| Competition-level mathematics | MATH | 2024 | 108.8 | o3-mini high |
| Competition-level mathematics | MATH | 2025 | 109 | GPT-5 (High Effort) |
| Multimodal understanding and reasoning | MMMU | 2023 | 71.9 | Gemini 1.0 Ultra |
| Multimodal understanding and reasoning | MMMU | 2024 | 94.7 | o1 |
| Multimodal understanding and reasoning | MMMU | 2025 | 106.1 | Gemini 3 Flash |
| PhD-level science questions | GPQA Diamond | 2023 | 47.8 | GPT-4 with search (backoff to CoT on abstention) |
| PhD-level science questions | GPQA Diamond | 2024 | 108 | o3 |
| PhD-level science questions | GPQA Diamond | 2025 | 114.5 | gemini-3-pro-preview |
| Agent multimodal computer use | OSWorld | 2024 | 5.7 | GPT-4 |
| Agent multimodal computer use | OSWorld | 2025 | 75.7 | Claude Opus 4.5 (no thinking) |
| Autonomous software engineering | SWE-bench Verified | 2024 | 64.3 | Claude 3.5 Sonnet / GPT-4o |
| Autonomous software engineering | SWE-bench Verified | 2025 | 98.7 | Claude Opus 4.5 |
| Mathematical reasoning | AIME | 2024 | 93.3 | OpenAI o1 |
| Mathematical reasoning | AIME | 2025 | 98.1 | Gemini 3.1 Pro Preview |
Benchmarks mirrored
Every item is original, written in the spirit of a published test
| Benchmark | Faculty | Items | AI result | What it means |
|---|---|---|---|---|
| RIAC | Attention | 7 | 51% | frontier models with one planted lure |
| ARC-AGI | Problem solving | 7 | 95.0% | best model on ARC-AGI-2 (human panel: 100%) |
| LearningBench | Learning | 6 | 0.85 | best model; 11 of 14 models under 0.50 |
| EphLangBench | Learning | 6 | 89% | best model; GPT-5.4 at 36% |
| MetaBound | Metacognition | 7 | 100% | of fake questions caught by the best models; Claude Opus 4.6 caught none |
| GAUGE | Metacognition | 5 | 0 folds | most accurate model never folded in 270 problems |
| Turn Bench | Executive functions | 4 | 80% | best re-adaptation rate; Opus 4.6 at 42% |
| HedgeDecode | Social cognition | 7 | ≤ 0.60 | every model on the hardest delivery task |
| MEDLEY-BENCH | Metacognition | 6 | split | some models weigh arguments, others count heads |
| ABC | Attention | 5 | weakest | on questions that hinge on which group an entry is in |
| Sustained attention to response task | Attention | 3 | n/a | no fair comparison: models don't tire or drift |
| Verbal fluency task | Generation | 3 | n/a | untimed listing is trivial for language models |
| Binz & Schulz (2023) | Learning | 3 | 2.72 | GPT-3's average regret on two-machine choices (people: 3.24; lower is better) |
| PlanBench | Executive functions | 3 | 97.8% | o1-preview on block-stacking plans (52.8% with disguised names) |
| Luo et al. (2025) | Executive functions | 3 | 93.1% | average on clashing colour words, 47 vision-language models |
| Strachan et al. (2024) | Social cognition | 3 | 4 of 5 | theory-of-mind tests where GPT-4 matched or beat people |
| Webb, Holyoak & Lu (2023) | Reasoning | 3 | 3 of 4 | verbal analogy sets where GPT-3 matched or beat people |
| Cognitive reflection test | Reasoning | 6 | 96% | GPT-4 on cognitive reflection tasks (people: 38%) |
| Huang et al. (2025) | Memory | 6 | 26% | best model keeping numbers in mind without writing them down |
| CUTE | Perception | 6 | 13.1% | best published score at swapping two letters in a word |
| ClockBench | Perception | 7 | 83.3% | best model (average person: 90.7%) |
| Constrained writing | Generation | 5 | n/a | no direct human-vs-AI benchmark yet |
| Wason selection task | Reasoning | 6 | n/a | no clean held-out AI measurement (people: < 10%) |
| SimpleBench | Problem solving | 7 | 81.9% | best model (people: 83.7%) |
| Hendrycks et al. (2025) | Memory | 5 | 0% | long-term memory across sessions |
Can you beat AI?
Seven items, about five minutes. Your answers stay anonymous.