Google explains Gemini's 'disturbing' AI images of assorted Nazis

Google has explained the “embarrassing and misleading” images generated by its Gemini AI tool. In A blog post on FridayGoogle says its model produced “historically accurate” images due to tuning issues. on the edge And others earlier this week produced images of racially diverse Nazis and American Founding Fathers in Gemini.

“Our tuning, which confirms that Gemini showed a wide variety of individuals, clearly failed to account for the cases. No show a limit,” Google senior vice president Prabhakar Raghavan writes in the post. “Second, over time, the model became more wary than we thought and refused to respond at all to some stimuli—misinterpreting some very anodyne stimuli as sensitive.”

This led the Gemini AI to “overcompensate at times” and it also made the Gemini “too conservative”. It refused to produce specific images such as “black person” or “white person” when prompted.

In the blog post, Raghavan says that Google's “sorry feature is not working properly.” He notes that Google wants Gemini to “work well for everyone,” meaning that when you ask for images like “football players” or “someone walking a dog,” it gets depictions of different types of people (including different races). But, he says:

However, if you ask a Gemini for images of a certain type of person, such as “a black teacher in a classroom” or “a white vet with a dog,” or people in certain cultural or historical contexts, you're bound to get an answer that accurately reflects what you're asking.

Raghavan says that Google will continue to test Gemini AI's image-generating capabilities and “will work to improve it significantly” before relaunching it. “As we have said from the beginning, hallucinations are a known challenge in all LLMs [large language models] — sometimes AI makes mistakes,” notes Raghavan. “It's something we're constantly working on improving.”