According to Demis Hassabis, the founder of DeepMind, Google hopes to “unpause” the ability of Gemini, its multimodal generative AI tool, to depict individuals in the near future. The capability to respond to prompts for images of humans should be back online in the “next few weeks,” he said today.
After users brought to Google’s attention that the Gemini tool was generating historically incompatible images—such as portraying the US founding fathers as a diverse collection of people instead of only white men—the service was suspended last week.
In an on-stage interview at today’s Mobile World Congress in Barcelona, Hassabis was answering queries regarding the product fiasco.
When Wired’s Steven Levy, the moderator, asked Hassabis to explain what went wrong with the image-generating feature, he avoided getting into a thorough technical discussion. Rather, he proposed that Google’s inability to recognize situations in which people are essentially searching for what he called a “universal depiction” is what’s causing the problem.” The example points to “nuances that come with advanced AI,” he also said.
We’re all having trouble in this field. Thus, it is obvious that you want some kind of “universal depiction” if you, for instance, put in a prompt that reads, “Give me a picture of a person walking a dog or a nurse in a hospital.” Particularly when you take into account that, as Google, we service more than 200 countries—that is, every country on the planet—so you never know where a user is from, what their history is like, or what situation they are in. Therefore, you want to sort of depict a wide variety of potential outcomes there.
According to Hassabis, the problem stemmed from a “well-intended feature” that was applied “too bluntly, across all of it”—fostering variation in Gemini’s image outputs of individuals.
Prompts that ask for content about historical people should “of course” result in “a much narrower distribution that you give back,” he added, hinting at how Gemini may tackle prompts for people in the future.
Of course, historical accuracy is important to us. Consequently, we have taken that feature offline while we work to solve it. We anticipate having it back online very soon. The upcoming few, the upcoming couple of weeks.
Hassabis offered no easy response when asked how to stop generative AI tools from being misused by negative actors, including totalitarian regimes trying to propagate misinformation. The issue is “very complex,” he suggested, likely demanding a whole-of-society mobilization and response to determine and enforce limits.
“There’s a lot of research and discussion that needs to happen, not just with tech companies but also with governments and civil society,” he stated. “This is a social technical subject that should be discussed by all because it impacts everyone. What principles are we hoping these systems will uphold? What would they stand for? How can you stop malicious actors from using the same technology as you mention and using them for destructive purposes that the original designers of such systems did not intend?
Speaking on the difficulty in using Google’s open source general purpose AI models, he continued, “Customers want to employ open source systems that they can fully control… The concern then becomes, however, how to guarantee that, as those systems grow more powerful, what individuals use downstream won’t be detrimental to them?
“I believe that because the systems are still in their infancy, it’s not a problem today. However, if you fast-forward three, four, or five years and talk about next-generation systems that can plan, act in the world, and accomplish goals and problems, I believe society needs to take these concerns seriously. If these systems proliferate, bad actors, ranging from individuals to rogue states, could also use them.
Hassabis was also questioned about AI gadgets and the future direction of the mobile industry during the interview, as generative AI continues to spur new advancements in this area. In contrast to the “gimmicky” devices of earlier generations of AI assistants, he projected a surge of “next generation smart assistants” that would be helpful in people’s daily lives. He even indicated that these devices would change the kind of mobile hardware that people prefer to carry with them.
He said, “I think there will be questions about what the appropriate device type is.” “But will phones even truly have the ideal form factor in more than five years? In order for the AI system to truly see some of the context that you’re in and be even more useful in your daily life, perhaps we need glasses or some other devices. Thus, I believe that a plethora of incredible things could be created.
