Human Baseline

Methodology

How the test works, where the numbers come from, and what it can and can’t say.

v0.1Data terms ↓References ↓

01 · Framework

Which faculties

The ten faculties come from Google DeepMind’s 2026 framework, which draws on psychology, neuroscience and cognitive science: eight building blocks (perception, generation, attention, learning, memory, reasoning, metacognition, executive functions) and two composites (problem solving, social cognition). The authors present the list as a hypothesis, not a final account of cognition.

02 · Items

Which tests

Every item is original, written to mirror a published benchmark: the winners of the 2026 Kaggle × Google DeepMind hackathon, ARC-AGI, ClockBench, SimpleBench, and classic psychology tasks (Wason selection, Wisconsin card sorting, artificial grammar learning).

Each item type has several versions and each participant gets a random one, so answers posted online spoil only one variant. Items are versioned; any wording change creates a new version and responses are always tied to the exact text shown.

03 · Scoring

How you’re scored

Partial credit where an item has several parts, averaged within each faculty. Once enough people have answered, item difficulty is estimated from the data (item response theory) and your profile is reported as percentiles against everyone tested.

04 · Comparison

Where the AI and human numbers come from

Today each AI marker is the best published result on the benchmark an item mirrors, mostly from 2025 and 2026, and several are reported by the labs themselves. Showing the best model flatters AI: on metacognition, two other models scored 0%.

Human references vary widely, from nine non-specialists on SimpleBench to PhD experts on GPQA. Two (attention, metacognition) are our estimates. That is the gap this project exists to close: the same items, answered by a large and described sample of people and by the models themselves.

05 · Limits

What this is not

It is not an IQ test or a clinical instrument. The short test has one item per faculty and no norms yet. Participants self-select, so the sample is not representative of any population; we collect background information so results can be reweighted.

06 · Your data

Data terms

  • You must be 18 or older to take part.
  • We record your answers, response times, and whether you pasted text or switched tabs during an item. We do not record keystrokes or anything outside this site.
  • The short test is anonymous: a random session ID, no name, no email, no IP address stored with your answers.
  • If you create an account, we store your email (for sign-in only) and link your past sittings in this browser to it. The short background questionnaire (age band, gender, country, language, education, work, relative income, AI use…) is optional question by question and stored in a separate table from your email.
  • The questionnaire asks whether you used outside help (AI, search, another person). Answers with help are kept but flagged; they never change your score.
  • A longer questionnaire (about 10 minutes) will come with the full battery. It may ask about sleep and health, and will request separate, explicit consent for that.
  • Anonymised, aggregated data may be published as an open research dataset. Combinations of background answers that could single someone out are coarsened or removed before release.
  • You can delete your account and every linked answer at any time from your account page.

07 · References

References

Take the test →