Google is working to bring back Gemini’s ability to create images of people after pausing the feature over historically inaccurate results. DeepMind founder Demis Hassabis said the capability could return in the “next few weeks,” while acknowledging that a diversity-focused feature had been applied too broadly.
Why Google Paused Gemini’s People Images
The issue emerged after users pointed out that Gemini was producing images that did not match historical context. One cited example involved the tool depicting the U.S. Founding Fathers as a diverse group of people, rather than only white men.
Google suspended the image-generation capability last week. The pause applies to Gemini’s ability to respond to prompts that ask for images of humans, not to the whole multimodal generative AI tool.
Hassabis addressed the problem during an onstage interview at the Mobile World Congress in Barcelona today. When asked by Wired’s Steven Levy what had gone wrong, he did not give a detailed technical breakdown. Instead, he framed the failure as a context problem: Gemini did not properly distinguish between broad, generic prompts and prompts that require historical precision.
The Difference Between Universal And Historical Prompts
Hassabis described some prompts as asking for a “universal depiction.” His examples included requests such as “give me a picture of a person walking a dog or a nurse in a hospital,” where a wide range of possible depictions may be appropriate.
That approach is shaped by Google’s global audience. Hassabis noted that Google serves “200+ countries,” meaning the company often does not know where a user is coming from, what background they have, or what context they bring to a prompt.
In those generic cases, he suggested, Gemini should present a broad range of possibilities. But the same logic becomes problematic when the subject is historical. A prompt involving historical people should not be treated the same way as a prompt for a general person or profession.
Hassabis said the problem came from a “well-intended feature” intended to increase diversity in Gemini’s images of people. The feature, he said, had been applied “too bluntly, across all of it.”
That distinction matters because generative AI systems often need to infer what a user means from very short prompts. A request for a nurse in a hospital leaves many visual choices open. A request involving a specific historical context leaves far fewer.
What The Fix Is Expected To Address
Hassabis indicated that Gemini may handle people-related prompts differently depending on the kind of subject requested. For historical people, he said the system should return “a much narrower distribution.”
He also emphasized Google’s stated priority: “We care, of course, about historical accuracy.” Google has taken the feature offline while it works on the fix, and Hassabis said the company hopes to restore it “in very short order. Next couple of weeks, next few weeks.”
The broader lesson is not only about one product mistake. It shows how AI image systems can struggle when a safety or representation goal is applied without enough attention to prompt context. A feature that may make sense for open-ended depictions can produce wrong results when accuracy is the main requirement.
For users, the episode highlights a basic tension in generative AI tools:
- Generic prompts may benefit from broader representation and varied outputs.
- Historical prompts require closer alignment with known context.
- Product safeguards can create new errors if they are not sensitive to the type of request.
The Misuse Question Is Bigger Than One Feature
The interview also moved beyond the Gemini image controversy. Hassabis was asked how generative AI tools can be kept from bad actors, including authoritarian regimes seeking to spread propaganda.
He said there was no simple answer and called the issue “very complex.” In his view, the response cannot come only from tech companies. He said “civil society and governments” also need to be part of the research and debate.
Hassabis described the problem as a “social technical question” involving the values these systems should have, what they should represent, and how to prevent harmful use by people who repurpose technology beyond its creators’ intentions.
He also discussed open source, general-purpose AI models, which Google offers as well. Customers, he said, want systems they can fully control. But that raises a downstream question: how to ensure increasingly powerful systems are not used in harmful ways.
Hassabis suggested the concern is limited today because the systems remain “relatively nascent.” But he warned that the stakes could change in “three, four or five years” as next generation systems gain planning capabilities, act in the world, and solve problems and goals.
AI Assistants Could Change Devices Too
Hassabis was also asked about AI devices and the mobile market. He predicted a wave of “next generation smart assistants” that are useful in everyday life, contrasting them with earlier AI assistants he described as “gimmicky.”
That could affect hardware choices. Hassabis questioned whether the phone will remain the ideal form factor in “five plus years’ time.” He suggested that glasses or other devices may help AI systems see more of a user’s context and become more useful in daily life.
For now, Google’s immediate challenge is narrower: restoring Gemini’s ability to generate people while avoiding the historical-image errors that led to the pause. The company’s fix will be watched as a test of whether major AI products can balance representation, accuracy and context without flattening very different kinds of prompts into the same output logic.