Google's DeepMind Establishes a New Unit Focused on AI Safety
Google, through its flagship GenAI model, Gemini, exhibits concerning capabilities. On prompt, it can produce misleading content about various topics, from elections to sports events or even scientific incidents like the Titan submersible implosion, complete with plausible but false citations. This has drawn criticism, especially from policymakers, highlighting the ease with which such AI tools can disseminate disinformation and deceive.
In response to these concerns, Google, amidst a recent reduction in its workforce, is shifting its focus towards AI safety. Today, Google DeepMind, known for spearheading projects like Gemini, announced the creation of a new entity: AI Safety and Alignment. This organization will consolidate existing teams working on AI safety while also integrating specialized groups of GenAI researchers and engineers.
Although Google has not disclosed specific hiring numbers resulting from this initiative, it mentioned the inclusion of a new team dedicated to ensuring safety around artificial general intelligence (AGI). This move aligns with efforts by OpenAI, Google's rival, which established a similar division last July to address similar challenges posed by advanced AI systems.
The decision to have two groups working on the same issue prompts speculation, given Google's limited transparency. However, it's notable that the new team within AI Safety and Alignment will be based in the U.S., closer to Google's headquarters. This reflects the company's proactive approach to keeping pace with AI advancements while demonstrating responsible AI deployment.
Anca Dragan, a former Waymo staff research scientist and UC Berkeley professor, will lead this new team. Her expertise in AI safety, despite being divided between DeepMind and academia, underscores the interdisciplinary nature of the field and the need for collaborative efforts.
The AI Safety and Alignment organization will prioritize several areas, including preventing the dissemination of erroneous medical advice, ensuring child safety, and mitigating biases in AI systems. Dragan emphasizes the importance of understanding human preferences and values, robustness against adversarial attacks, and incorporating diverse viewpoints into AI development.
However, skepticism surrounding GenAI tools persists, especially concerning deepfakes and misinformation. With public concern about the proliferation of misleading content, there's pressure on tech companies to address these issues effectively.
Dragan acknowledges the complexity of AI safety challenges and commits DeepMind to investing more resources in this area. She emphasizes the need for frameworks to evaluate GenAI model safety risks and proposes strategies like incorporating uncertainty estimates and implementing inference-time monitoring.
Despite these efforts, questions remain about the reliability of AI systems and their potential impact. As Dragan notes, ensuring the safety and effectiveness of AI models requires ongoing vigilance and collaboration across various stakeholders.
Ultimately, the success of these endeavors will depend on how effectively AI systems can adapt to evolving challenges while minimizing harm to users and society at large.
