AI is now rewriting the past with fake vintage photos. Learn the dated-source test that separates a real archival photograph from an AI-generated one, then prove it by playing guessNow, where every reveal shows the photographer, licence and source.
The AI-generated historical photo is no longer a fringe curiosity. York University researchers set out the problem in March 2026: AI reconstructions introduce historical inaccuracy, raise ethics questions, and bypass the primary sources historians rely on. The Halifax Examiner reported in July 2026 on a viral fake of a '1922 Lunenburg schoolhouse fire' spreading across genealogy groups with every tell of an AI image. AAP FactCheck documented Holocaust-survivor photographs being repurposed into AI clickbait, and France24 and AFP covered how 'AI slop' is distorting public memory of the Holocaust. The wave is global: in June 2026 Lead Stories tracked a photo falsely dated 1903, captioned 'Bengali Woman Carrying A British Merchant', straight back to OpenAI image tools via the image's SynthID watermark; DFRAC in India showed a widely-shared picture of Jawaharlal Nehru with a woman grabbing his collar was AI-generated at 99.9 percent confidence despite a real verbal account of a similar incident; and Greece Fact Check confirmed a viral image of the Acropolis dwarfed by a new skyscraper was fabricated. In every case the pattern is the same: a plausible-looking vintage frame presented as fact, with no archive record behind it. This is not a footnote of the year - it is an active news cycle about the past itself being rewritten.
Once you know what is being faked, the images stop being either 'obviously fake' or 'indistinguishable'. The useful division is by what the generator did: an invented scene that never happened; a reconstruction trying to resurrect a real event convincingly; a generative restoration that fills in missing parts of a real old photo; an AI colorisation that recolours a genuine black-and-white original; and a real photograph carrying a false caption or a wrong date. Each deserves a different check - and only one of the five, the true-photo-with-false-caption case, can fool you on pixels alone.
The mistake most people make is to stare at the grain, the hands, the eyes and ask 'does this look old enough?'. Generators are now good enough that this fails. The reliable question is provenance: can you trace WHO made this image, WHAT it shows, WHERE and WHEN, back to a catalogued record - a library, an archive, a museum, a press photo, a named photographer with a dated negative? A real archival photograph has a paper trail; an AI fake is born with none. Genuine archives like the Library of Congress, the National Archives and state collections index their photographs to a call number and a dated record, and that is the check that works at scale.
guessNow is the only game in this lane that treats provenance as the point of play: every reveal shows the photograph's photographer, its licence and its source, straight from the archive record. Open a round and you are dropped into a real archival frame, revealed piece by piece; you commit to a year; and then the dated record appears to confirm or correct you. The daily challenge gives everyone the same five real moments a day, and the archive holds the full dated catalogue. In a game like guessNow, the dated frame itself is the point - the habit of asking 'who made this, and where is the record?' is exactly the anti-AI-slop literacy the news cycle is calling for.
Trace the source first: is there a named photographer, a dated negative or a catalogued collection behind it? Reverse-search a plausible-known version and compare; an AI image usually has no earlier version to find. Ask whether a plausible real event, place and date are attached, and whether an archive you could name would hold it. Then still test yourself: play a round of guessNow with a real archive photograph and read the photographer, licence and source on the reveal. Teachers may want the primary-source activity, which turns the check into a ten-minute lesson.