Methodology

The labelling principle

Every synthetic item in OneIsFake is labelled at the moment it is created, not classified afterwards. We generate it, so we know what it is. That single decision removes the biggest source of error in AI-detection material: nobody on our side is squinting at an image and guessing, and no automated detector is being trusted to produce ground truth.

Real items are only used when their provenance is documented — an openly licensed photograph with a traceable source, not an image scraped from a feed where the origin is unknowable. If we cannot establish that a photograph was captured by a camera, it does not become a "real" answer.

Where the media comes from

• Real: openly licensed photography from sources that publish licence and attribution information, plus material we have produced ourselves.
• Synthetic: generated with current, publicly available image and video models across several vendors, using prompts written to imitate ordinary photography rather than obvious "AI art".
• Text and news items: written by models for the synthetic side and drawn from published reporting for the real side, with the source recorded.

Prompting for realism matters more than it sounds. Test sets built from default, over-stylised generations are easy, and they teach a tell — glossy lighting — that fails immediately on output someone actually intends to deceive with.

Balance and difficulty

Decks are balanced so that the split between real and synthetic is close to even and not predictable from position or order. A player who always answers "AI" should land near chance, not above it. Rounds mix subjects — faces, scenes, streets, studio and candid material — because detection accuracy varies sharply by subject, and a face-only test overstates how good you are.

We keep older model generations in rotation on purpose. Detection skill trained on one generation transfers poorly to the next, so a set consisting only of the newest release would measure familiarity with that release rather than a general eye.

Which generations are represented

The vendors and model families in rotation, era by era, are listed on models and generations. The tells that each generation tends to leave behind — and the ones that have been patched out — are on how to spot AI, graded by how well each one is actually supported.

What we deliberately exclude

• Real people depicted in synthetic media without consent, including public figures. Face swaps of identifiable individuals are not used as puzzle content.
• Sexual content, gore, and material depicting minors, in either the real or synthetic half.
• Political disinformation reproduced as-is. Where a mode teaches fake-news detection, the synthetic items are invented rather than recycled real falsehoods, so playing the game never spreads a live claim.
• Watermark or metadata puzzles disguised as perception tests. If an item is only solvable by reading an embedded provenance signal, it belongs in an explainer, not a round.

What the score means, and what it does not

Your accuracy in OneIsFake measures how well you read this set, under time pressure, on a screen. It is not a certification, and it does not transfer to a claim about any specific image you encounter elsewhere. Nothing on this site authenticates media: a tell shifts the odds, while provenance data and locating the original source are the only things that come close to proof.

Corrections

Mistakes happen: a mislabelled item, a real photograph whose provenance turns out to be weaker than we thought, a pair that is unfairly hard for a reason unrelated to the tell. When that is reported, we remove the item first and investigate afterwards, and we correct the explainer pages when the evidence for a tell changes.

Report anything that looks wrong through contact. Include the round or image if you can — it makes the item findable.

Who maintains this

OneIsFake is designed and built by Waybound.ai, which also maintains the explainer pages referenced above. Background on the team and the reason the project exists is on the about page.