Google Pauses Gemini People Image Generation After Accuracy Failures
Google temporarily halts Gemini's ability to generate images of people after the feature produces inaccurate and offensive results due to overcompensating safety tuning.
Google pauses the image generation of people in its Gemini conversational app after users report inaccurate and offensive results. The feature, which launches just three weeks ago, relies on an underlying AI model called Imagen 2. Google admits the feature misses the mark and apologizes to users while it works on an improved version.
The issues stem from the company's attempt to make the AI show a diverse range of people for general prompts, such as requests for pictures of football players or dog walkers. However, the safety tuning fails to account for specific historical or cultural contexts where a range of people is inappropriate. Additionally, the model becomes overly cautious and wrongly interprets harmless prompts as sensitive.
These two miscalculations cause the AI to overcompensate in some situations and refuse to answer others entirely. Google emphasizes that the Gemini app is a distinct product from its core search engine and underlying AI models. The company keeps the people image generation feature paused until it resolves these tuning problems.