I just saw a Dutch-speaking user who was trying to look up more information about a photo of an archaeological dig (a black and white photo of a man and a woman) and Gemini interjected, in Dutch, that it was a photo of a previous Dutch royal couple. There is no particularly strong resemblance.

I think this is evidence that the language Gemini is told to use (from the user’s browser settings) has a massively disproportionate influence on the factual aspect of its answers, as if the fact that a Dutch user is searching for the source of a photograph implies it’s a photo of Dutch people and only Dutch personages need be considered. I’m willing to bet that the smaller the language, the more extreme this type of hallucination becomes.

One more peril to be aware of, and one more way in which non-English-speaking people are systematically neglected by technology

for filters: AI LLM