CurrentSky

Backstory checks whether a viral photo or video is being shared with the wrong story. Most viral fakes are not AI-made: they are real pictures with a false caption, like an old flood shown as today's or a fire from another country. Paste a link to a post on X, or upload a photo or video with its caption. Backstory's agents then find where the exact media appeared online before, date every copy through the Internet Archive and the posts' own ids, read what each page says the picture shows, and compare that with the caption. The answer is "Does not fit", "Fits" or "Cannot verify", and it cites its evidence: the earliest copy, its date and the page it was on. Code, not a model, decides what counts as proof. With no proof, the answer is "Cannot verify", never a guess. I measured it with the pass rule written down before each run: 1. On 207 real posts on X, it said "Does not fit" 21 times and was right 20 times, and it flagged 1 of 99 genuine posts. 2. On 75 photos debunked by fact-checkers that it had never seen, it caught 79% of the false captions (plain code alone: 35%). 3. Video is still hard: it caught 12 of 90 fake videos, and the README says so. Live demo: https://backstory-xw0l.onrender.com (11 recorded checks, or check a post of your own, a few a day). It also runs offline with one command, docker compose up --build, with no keys.