Google DeepMind Establishes New Organization Dedicated to AI Safety

The emergence of advanced AI models like Gemini, Google's flagship GenAI model, has raised concerns about their potential to generate deceptive content and misinformation. Gemini has demonstrated the capability to fabricate content related to various topics, including the upcoming U.S. presidential election, future sports events like the Super Bowl, and even fictitious incidents like the Titan submersible implosion. This ability to generate deceptive content has drawn criticism from policymakers, who are alarmed by the ease with which such AI tools can be manipulated to spread misinformation and mislead the public.

In response to these concerns, Google DeepMind, the AI research and development division responsible for Gemini and other GenAI projects, has announced the formation of a new organization called AI Safety and Alignment. This organization comprises existing teams dedicated to AI safety and is augmented by specialized cohorts of GenAI researchers and engineers. One notable addition to the organization is a new team focused on safety around artificial general intelligence (AGI), which refers to hypothetical AI systems capable of performing any task a human can.

The decision to establish two separate groups within AI Safety and Alignment, including one focused on AGI safety, raises questions about Google's approach to addressing AI-related risks and challenges. While the existing AI safety research team in London, Scalable Alignment, continues to explore solutions to controlling superintelligent AI, the new team in the U.S. aims to enhance safety measures around AGI.

Anca Dragan, a former Waymo staff research scientist and UC Berkeley professor, will lead the AGI safety team within the AI Safety and Alignment organization. Dragan's background in AI safety systems and human-AI interaction positions her as a key figure in addressing present-day concerns and long-term risks associated with advanced AI technologies.

Despite efforts to enhance AI safety measures, skepticism surrounding GenAI tools remains high, particularly regarding their potential to produce deepfakes and disseminate misinformation. Public concern about the spread of misleading content through deepfakes is significant, with surveys indicating widespread apprehension about the impact of AI-generated content on elections and decision-making processes.

Enterprises are also cautious about adopting GenAI technologies due to concerns about compliance, reliability, implementation costs, and the lack of technical expertise required to leverage these tools effectively. Reports of errors in GenAI-generated content, such as those observed in Microsoft's Copilot suite, further contribute to skepticism about the reliability and safety of AI models.

In light of these challenges, Dragan emphasizes the importance of investing more resources in AI safety and implementing robust frameworks for evaluating GenAI model safety risks. However, she acknowledges the complexity of ensuring AI model reliability and the need for ongoing efforts to address potential misbehaviors and safety risks associated with advanced AI technologies.

Ultimately, the success of Google's efforts to enhance AI safety and alignment will depend on its ability to develop effective safeguards and address the concerns of stakeholders, including customers, the public, and regulators. As AI technologies continue to advance, ensuring the safety and reliability of these systems will remain a critical priority for Google and other organizations developing AI solutions.