About The Tell

The Tell is a calibrated perception instrument disguised as a game. You see pairs of creative works made to the same brief, with no names and no production labels. You choose what you prefer, guess how each was made, and say how sure you are. The reveal returns three numbers people habitually conflate, alongside the makers' own account of their process.

The three scores

• Sensitivity (d′): how well you can tell the methods apart, separate from which way you lean. 0 is chance. It is computed from your accuracy with a correction so a perfect run is finite.
• Bias (criterion c): whether you call things “AI” or “human” regardless of the truth. It needs single-item rounds with an unannounced base rate, so it appears from edition one, not in the demo.
• Calibration (Brier score): whether your confidence matches your accuracy. 0 is perfect; saying 50% every time scores 0.25.

Preference is asked first and scored separately, with a Bradley-Terry model (which estimates each work's latent quality from pairwise choices) once enough judges have played. Every score carries an interval. Only your first attempt is scored. Fewer than 12 scored pairs yields no certificate.

What this does not claim

Judges self-select and the works are volunteer entries made on best effort, so published accuracy describes this sample, not social feeds or the population. Provenance is self-declared by makers and spot-checked, and every chart says so. Regions are never ranked on quality.

What creators get, and what we promise them

Creators enter hand-made and AI-assisted work on the same brief and compete for an equal prize on each side. Each gets a credited reveal, a private feedback report and a provenance record. Works are displayed only, never licensed to anyone, and never used for training. Creators can withdraw at any time. The platform is neutral on method and partisan on craft and honesty.

What we do with your judgements

Judgements (your choices, confidence, response time and any rationale you type) are logged anonymously for research: fitting the preference and item models above, and publishing aggregate results with their caveats. No account is needed; a cookie links your sessions so repeat attempts are not scored. Any dataset release contains judgements and metadata only; works are referenced by URL and never included.

The name

A tell is the thing people claim to see. The name was chosen blind from a shortlist written half by a person and half by a model; which half it came from has not been revealed. You can vote on it. We will not answer.

Contact

Questions, concerns or a withdrawal request: contact details are published with edition one.