Google paused Gemini's people images after a tuning failure

A launch, then a pause

Google introduced its Gemini model family on 6 December 2023, describing it in the launch post as its "most capable and general model yet", built in three sizes and reported to exceed prior results on most of a set of 32 academic benchmarks it cites. Weeks after Gemini's image-generation feature for people reached users, Google disabled it. In a post dated 23 February 2024, the company states the feature had produced "inaccurate historical" and otherwise "embarrassing" depictions of people, and that it had "turned the image generation of people off" while it worked to "improve it significantly before turning it back on".

What Google says caused it

The February post gives two stated causes. First, tuning intended to ensure Gemini "showed a range of people" had, in the company's own words, "failed to account for cases that should clearly not show a range" - producing anachronistic or historically inaccurate results when a prompt called for a specific, non-diverse depiction. Second, the model became overly cautious in a separate direction, wrongly refusing to answer certain prompts entirely that had no bearing on the first problem. Google frames both as consequences of the same underlying tuning pass, not two unrelated bugs, and commits to "extensive testing" before restoring the feature.

A single vendor's account of its own failure

This is a company's explanation of its own product, offered without an independent technical audit available to this analysis, and it should be read as a vendor's account rather than a verified root-cause finding. What it does establish, on its own terms, is a general pattern worth carrying beyond this one incident: a mitigation aimed at one failure mode - unrepresentative outputs in one direction - can introduce a different failure mode - inaccurate or overcautious outputs in another - when the underlying tuning is applied broadly rather than tested against the specific cases it needs to handle.

• When a system is tuned to correct one class of output, what tests were run for cases where that correction is inappropriate?
• Was the fix tested against historical or context-specific prompts before wider release, or only against the cases the original problem was named for?
• What did the vendor change when the feature was restored, and is that change documented anywhere a reader can check?

The episode is a dated example of a general risk in tuning generative systems for demographic representation: correcting a narrow, visible failure can shift the same underlying mechanism into a different, equally visible failure. Google's own account names the mechanism; it does not, on this evidence alone, prove the company has since resolved it.