Google aims to unpause Gemini's people generation in a few weeks

Ask an AI to generate the American Founding Fathers, and it gives you a diverse cast that defies historical reality. After days of mockery and intense backlash, Google abruptly turned off Gemini's human image generation last week.

DeepMind co-founder Demis Hassabis took the stage at MWC in Barcelona today to explain what happened and announced that human image generation should return in the next few weeks.

A blunt fix gone wrong

When pressed on stage by Wired's Steven Levy, Hassabis bypassed the dense technical weeds and admitted Google applied a well-intended feature far too bluntly across the board.

He explained that Google serves over 200 countries, meaning generic prompts like a person walking a dog or a hospital nurse genuinely need a universal depiction reflecting varied backgrounds.

However, the model failed to recognize when a prompt demanded historical accuracy rather than broad demographic variety. Hassabis noted that requests for historical figures should naturally produce a much narrower distribution of outputs.

Google pulled the feature to patch that nuance, and Hassabis claims the fix will land in short order, likely within the next couple of weeks.

Beyond the image fiasco, Hassabis warned about larger risks looming in AI. When asked about bad actors repurposing generative models for state-backed propaganda, he admitted tech giants cannot solve it in isolation.

He also touched on open-source AI, cautioning that while models today are relatively nascent, next-generation systems with real planning capabilities could pose massive dangers in three to five years if rogue states get full control.

On physical hardware, Hassabis doubted smartphones will remain the dominant AI form factor in five years, suggesting contextual smart glasses might finally replace today's gimmicky digital assistants.

Why it matters

This snafu shows just how clumsy AI alignment still is when forced through crude systemic filters. Google tried to hardcode diversity and accidentally broke historical reality, triggering an ideological firestorm that damaged trust in its flagship model.

For creators and professionals, this means AI image generators remain unpredictable when handling delicate historical or cultural prompts. Relying on them for accurate context right now is risky without human verification.

We think this incident forces a major rethink across the industry. Slapping brute-force diversity prompts on top of user queries does not work. Moving forward, models will need genuine contextual reasoning about when to embrace variety and when to stick strictly to historical facts.