Know what the
numbers mean.
The question, the audience and the collection window matter as much as the percentage.
Who takes part
Voluntary, opt-in responses from Superpower Daily readers; not representative of the population. One counted response per account, question and UTC collection month. First completed attempts only for scored tasks; practice attempts are excluded. Percentages include every answer, including uncertainty. Live percentages and subgroup results require at least 20 valid responses. Identity and individual answers are private. No population margin of error is claimed.
Five formats, different signals
Polls capture opinions. Experience checks capture self-reported outcomes, not verified behavior. Quizzes score defined questions. Blind comparisons record preferences between disclosed materials with randomized anonymous options; they are not universal model rankings. Benchmarks measure performance on a specified task, with the conditions shown before starting.
Collection and versions
Collection months use UTC, from midnight on the first day to midnight on the next month's first day. A recurring question opens a new window each month. A rotating question collects for its launch month. Material edits create a new question version. Existing monthly windows retain their original wording. Returning participant results only compare the same version across adjacent months.
Results and small groups
Live percentages unlock after a valid response and 20 valid respondents. Historical editions are public. A role or usage breakdown is published only when every nonempty group in that dimension has at least 20 responses, preventing small-group inference. Uncertainty and “Prefer not to answer” choices remain in their denominators. Percentages are rounded to one decimal place.
Attempts and measured performance
Scored task sessions last 30 minutes. Solutions and blind identities are revealed after submission. The first completed task per account per month contributes to results. Later attempts are explicitly marked practice. Recorded duration includes reading time and network latency, so it is not a controlled laboratory measurement. Familiarity with fixed tasks can improve later scores; month-to-month changes may include practice effects.
Privacy and review
Your account links repeat participation. Optional role and AI usage are stored with your response. Public exports contain aggregates only, never individual answers, account identifiers or contact details. Editorial exclusions need a recorded reason. Published reports and snapshots are immutable; corrections produce a new edition with a visible note.
State of AI reports
Quarterly and annual reports are assembled from published monthly editions. Months and question versions stay separate. Counts are not added together as if each month's participants were different people. Coverage can be partial; reports list exactly which months and questions they include.
Downloads, citations and embeds
Aggregated AI Pulse datasets and charts are available under CC BY 4.0. Attribute Superpower Daily, the exact question, collection month, version and edition, and link back. This license applies to Pulse aggregates and charts, not other site content or third-party test materials. Frozen iframe embeds always reference one published edition.
Each frozen JSON dataset has a SHA-256 fingerprint. To verify it, omit the root hash field, sort object keys recursively while keeping arrays in order, serialize compact JSON as UTF-8, and compare its digest with the published hash. This verifies the dataset, rather than the bytes of a PDF or CSV.
Survey standards
We follow disclosure principles described by AAPOR and stable-question principles described by Pew Research Center. Neither organization endorses this project.