The girl who wasn't there
A friend of the technology presenter Spencer Kelly scanned some old family photographs and asked Gemini to colour one. Ten seconds later it looked wonderful: natural skin tones, a sensible guess at the clothes. It had also decorated the room for Christmas, added a clock to the wall, and put a new person in the middle of the picture. "Who the hell is this?" he asks. His other theory, offered as a joke: maybe the little girl was there all along.
In favour- Colourisation and restoration are genuinely useful, and the result here was good enough that the additions were easy to miss, which is exactly why the example is worth keeping.
- The explanation is simple and generalises: the model saw an empty space and filled it with something statistically likely.
- It is the same move a language model makes with text. Ask about a paper that does not exist and it knows precisely what a citation looks like, authors, year, journal, pages. It has learned the costume. Confidence about the next token is not confidence about the world.
- A missing citation and a missing person are the same gap, filled the same way. Only the medium changes. Grounding with retrieval and tools reduces it; it does not remove it.
Our takeFor anyone digitising memory, archives, museums, local history groups, a box of prints at home: an AI-restored image is not a better copy of the original, it is a new image that looks like one. Keep the untouched master, label the derivative as AI-colourised in the file name and the catalogue record, compare side by side before sharing, and never treat it as evidence of who was in the room. If such an image is published and could pass as authentic, EU transparency rules for manipulated images apply. Not legal advice; the practical rule is simpler: when a machine fills a gap, write down that it did.
Source: «This is why you shouldn't trust generative AI, in one photo» — Spencer Kelly (TikTok, 22/9/2026)
← All Focus posts