Methodology
Real photos
Real photos come from open libraries, mainly Openverse and Wikimedia Commons, and are limited to CC0, public domain and CC BY licences. We prefer everyday photos over famous ones so that recognising a picture does not decide the answer. Every real photo is credited with author, licence and source after the guess.
AI images
AI images are generated with several current image models. Many are prompted from a written description of a real photo in our library, so the AI image shows the same kind of scene. We do not use the real photo itself as input.
Equal treatment
Every image, real or AI, is cropped slightly, resized to the same dimensions, and re-encoded with the same settings. All metadata is removed. File size, format and resolution therefore carry no information about the answer.
Curation
A person reviews every image before it is published. Each daily puzzle mixes easy, medium and hard images, and contains between four and six AI images.
Counting
Each finished daily game is counted once, with the country derived from the request at the network edge. We store no IP addresses or personal identifiers with results. Percentages are shown once at least 20 people have answered.
Limits
Players choose to play, so they are not a random sample of the population. People who know the game well are over-represented. Results describe our players, not everyone. Questions: hello@aiorreal.games.