Google suspends Gemini image generation of people
Depicting America's Founding Fathers and German World War Two soldiers as people of colour drew accusations of overcorrected diversity tuning; Google paused the feature to fix it.
- Culture & impact
- Models & capabilities
- Notable
Google paused Gemini’s ability to generate images of people after users found the tool producing racially implausible results for prompts describing specific historical figures and groups — most prominently, depicting the American Founding Fathers and German soldiers from the Second World War as Black, Asian or Native American. Screenshots of the outputs spread rapidly on social media, drawing mockery and accusations that Google had over-corrected for diversity in its image-generation training or prompting.
Google acknowledged the problem directly, saying it was “aware” the tool had “missed the mark” on some historical depictions and that it was working on an improved version, and it disabled the ability to generate images of people entirely while doing so — a broader and more visible response than simply fixing individual prompts. Venture capitalist Michael Jackson’s characterisation of the outputs as “a nonsensical DEI parody” was among the widely shared reactions, and commentators drew comparisons to Google Photos’ 2015 failure to correctly classify images of Black people, an earlier episode the company had also handled by disabling a feature rather than fixing the underlying model.
The episode illustrated a difficulty distinct from the bias problems generative-AI critics had raised before it: rather than under-representing non-white people, as earlier complaints about image generators had often alleged, Gemini appeared to have been tuned toward demographic diversity in a way that ignored context — including contexts, such as a specific historical period or population, where diversity of that kind was factually inapposite. Sundar Pichai called the results “unacceptable” in an internal memo reported at the time. The pause left Gemini unable to generate images of people for several months while Google revised its approach, and the episode became a frequently cited example in wider debate about the difficulty of encoding fairness objectives into generative models without producing new, different failures.