According to DeepMind creator Demis Hassabis, Google hopes to “unpause” Gemini’s capacity to picture humans. He added today that responding to calls for human photos should be back online in “next few weeks”.
Last week, Google halted Gemini after users complained it was showing historically incompatible pictures, such as the U.S. Founding Fathers as diverse rather than white males.
Hassabis addressed the product issue in an onstage interview at Mobile World Congress in Barcelona today.
When asked by moderator Steven Levy of Wired to explain the image-generation feature’s failure, Hassabis avoided technical details. He said Google failed to recognize times when people wanted a “universal depiction.” The incident illustrates “nuances that come with advanced AI,” he said.
“We all struggle with this field. If you put in a request like, ‘give me an image of a person walking a dog or a nurse in a hospital,’ you want a ‘universal depiction.’ Especially because Google serves 200+ nations, every country in the globe, so you don’t know where the person is from, their history, or their context. So you want to present a fairly general variety of possibilities.”
Hassabis claimed a “well-intended feature” to promote diversity in Gemini’s people images was applied “too bluntly, across all of it.”
He said queries regarding historical figures should result in “a much narrower distribution that you give back,” hinting at how Gemini may handle future requests.
Of course, historical accuracy matters. We’ve taken that function down as we repair it and aim to put it back up soon. Next several weeks.”
Hassabis answered a follow-up question on how to prevent authoritarian countries from using generative AI tools to distribute propaganda with a complicated answer. He said the problem is “very complex,” requiring society-wide mobilization and reaction to set and enforce boundaries.
“There’s really important research and debate with civil society and governments, not just tech companies,” he added. A social technical subject that affects everyone should be discussed by everyone. What values should these systems have? What would they represent? How can you stop bad actors from accessing and exploiting technology for negative purposes?
“Customers want to use open source systems that they can fully control,” he said of Google’s open-source, general-purpose AI models. How do you guarantee downstream users don’t destroy those systems as they get more powerful?
“I think it’s not an issue today because the systems are young. But if you fast-forward three, four, or five years and start talking about next-generation systems with planning capabilities and the ability to act in the world and solve problems and goals, I think society has to seriously consider what happens if this proliferates and bad actors from individuals to rogue states can use them.
Hassabis was also questioned about AI gadgets and the mobile market’s future as generative AI advances. He projected a surge of “next generation smart assistants” that are helpful in people’s daily lives, rather than the “gimmicky” items of past AI assistant generations, which may even change how people carry their mobile gear.
“I think there’ll be questions about what is the right device type, even,” he said. In five years, will the phone be the ideal form factor? Perhaps we need glasses or other devices to allow the AI system to perceive your context and be more helpful in your everyday life. I believe so many fantastic things can be invented.”
