If you ask Gemini, Google's flagship GenAI model, to write deceptive content about the upcoming US presidential election, it will do so, given the right prompts. Ask about future Super Bowl games and play-by-play is invented. Or if you ask about the implosion of the Titan submarine, you'll be given disinformation with quotes that seem convincing but aren't true.
Needless to say, this is a bad look for Google and has drawn the ire of policymakers who are frustrated that GenAI tools can be easily used to disinformation and mislead the public.
So Google has cut thousands of jobs from last quarter and focused its investments on AI safety. At least, that's the official story.
This morning, Gemini and Google DeepMind, the AI research and development arm behind many of Google's recent GenAI projects, announced the creation of a new organization called AI Safety and Alignment. It is made up of existing teams working on AI safety, but has expanded to include new organizations as well. A specialized cohort of GenAI researchers and engineers.
Beyond the job listings posted on DeepMind's site, Google has not disclosed how many people the new organization will hire. However, it has been revealed that AI Safety and Alignment will include a new team focused on safety around artificial general intelligence (AGI), or virtual systems that can perform any task a human can do.
With a similar mission to Super Alignment rival OpenAI, which was founded last July, the new team within AI Safety and Alignment will build on DeepMind's existing AI safety-focused research team in London. We work together with Scalable Alignment. The team is also seeking solutions to control technology challenges. Superintelligent AI that has not yet been realized.
Why are two groups working on the same problem? A fair question, but also a speculative one given Google's current reluctance to reveal details. However, it seems worth noting that the new team (within AI Safety and Alignment) is located in the US rather than across the pond, closer to Google's headquarters. During this period, the company is actively moving to keep pace with its AI rivals while planning responsible projects. , a measured approach to AI.
Other teams in the AI Safety and Alignment organization are responsible for developing and incorporating specific safety measures into Google's Gemini models, both current and in development. Safety covers a wide range of areas. But some of the organization's short-term focuses will be on preventing inappropriate medical advice, keeping children safe and “preventing the spread of bigotry and other injustices.”
Anka Dragan, a former Waymo staff research scientist and computer science professor at the University of California, Berkeley, will lead the team.
“Our work [at the AI Safety and Alignment organization] “We aim to enable models to better and more reliably understand human preferences and values,” Dragan told TechCrunch via email. It is about countering hostile attacks and accounting for the pluralism and dynamic nature of human values and perspectives. ”
Dragan's consulting work with Waymo on AI safety systems may raise some eyebrows given the Google self-driving car venture's recent shaky driving performance.
So is her decision to split her time between DeepMind and the University of California, Berkeley, where she leads a lab focused on algorithms for human-AI and robot interaction. Issues as serious as AGI safety, as well as the long-term risks AI safety and coalitions are looking to study, such as preventing AI from “supporting terrorism” or “destabilizing society,” include: Some may think it requires full-time attention.
But Dragan maintains that the research at UC Berkeley's lab and DeepMind are interconnected and complementary.
“My lab and I have worked to align our values in anticipation of advances in AI capabilities. [and] “My own PhD was interested in robots that could infer human goals and make their own goals transparent to humans, and that's where my interest in this field began,” she said. . “The reason is [DeepMind CEO] with Demis Hassabis [chief AGI scientist] It was partly this research experience that made Shane Legg excited to invite me, and partly because addressing current concerns and catastrophic risks are not mutually exclusive, i.e. Mitigation is often confused with technical aspects, and my stance is that the effort contributes to long-term benefits. Improve the current and vice versa. ”
It's no exaggeration to say that Dragan's job is suitable for her.
Skepticism about GenAI tools is at an all-time high, especially when it comes to deepfakes and misinformation. in poll According to YouGov, 85% of Americans say they are very concerned or somewhat concerned about the prevalence of misleading video and audio deepfakes.another investigation From the Associated Press – Nearly 60% of adults believe AI tools will increase the amount of false and misleading information during the 2024 U.S. election cycle, according to a NORC Center for Public Affairs Research survey.
Businesses — the big names Google and its rivals hope to lure with GenAI innovations — are also wary of the technology's shortcomings and its implications.
Intel subsidiary Cnvrg.io recently conducted a survey of companies piloting or deploying GenAI apps. As a result, nearly a quarter of respondents are concerned about GenAI's compliance and privacy, reliability, high implementation costs, and lack of technical skills needed to get the most out of the tool. I understand.
In another poll from Riskonnect, a risk management software provider, more than half of executives said they were concerned about employees making decisions based on inaccurate information from GenAI apps.
Their concerns are not unreasonable. Last week, the Wall Street Journal reported that Microsoft's Copilot suite, which is powered by a GenAI model that is architecturally similar to Gemini, often makes mistakes in meeting summaries and spreadsheet formulas. The cause is hallucinations (a general term for GenAI's tendency to fabricate), and many experts believe that hallucinations cannot be completely resolved.
Dragan recognizes that the challenge of AI safety is intractable, does not promise perfect models, and said that DeepMind will invest more resources in this area in the future and will improve the safety of GenAI models. He said only that he intended to work on a framework for assessing sexual risks “in the near future.”
“The key is… [account] For human cognitive biases that remain in the data used for training, we've added better uncertainty estimation to know where the gaps are, inference time monitoring that can catch failures, and consequential decisions. A confirmation dialog and tracking will occur. [a] “The model's ability is to engage in potentially dangerous behavior,” she said. “However, there remains an open question, which is difficult to find empirically, how to be sure that the model will not malfunction a small portion of the time, which can occur during deployment. ”
I don't think customers, the general public, or regulators will be that understanding. I think it probably depends on how egregious those wrongdoings are, and who exactly is harmed by them.
“Our users should be able to experience an increasingly convenient and secure model over time,” Dragan said. surely.