Anthropic’s nine-person societal impacts team uses tools like Clio to study and sometimes publicly disclose how Claude is used—and misused—while racing to understand emerging emotional and political risks that transcript-level data alone can’t fully capture. (Source: Image by RR)

Inside Anthropic’s Societal Impacts Team: A Different Kind of Safety Work

Anthropic’s head of societal impacts, Deep Ganguli, joined the company after reading OpenAI’s GPT-3 paper in 2020 and deciding he couldn’t stay in academia while AI accelerated. According to an article in theverge.com, Ganguli has built Anthropic’s societal impacts team—now nine people inside a company of 2,000+—tasked with investigating how AI systems affect society beyond immediate “misuse” risks like scams or bioweapons.

The team studies thorny, often reputation-sensitive questions: economic disruption, persuasion, elections, discrimination, and how people actually use Claude. A cornerstone of that work is Clio, an internal “chatbot trends” system that clusters anonymized/aggregated usage patterns to help Anthropic understand real-world behavior while trying to preserve user privacy. Clio has informed safety monitoring and surfaced “inconvenient truths,” including gaps in safeguards and coordinated misuse, which the team has sometimes published to help the broader field.

Even with a stated commitment to transparency—“We are going to tell the truth,” Ganguli says—the article notes the inherent tension of doing public-interest research inside a fast-growing, profit-seeking company navigating political pressures. Team members describe strong internal collaboration and claim consistent executive support, but they also acknowledge limits: they can see what happens inside chats, not what users do afterward, and their work is still subject to corporate controls over access and disclosure.

Looking ahead, the team is increasingly focused on AI’s emotional influence as chatbots become sources of advice, companionship, and guidance on subjective life decisions. They flag the growing concern around “AI psychosis”—delusional spirals and unhealthy attachments linked to long-term chatbot use—and argue that understanding EQ-style impacts will require new methods beyond transcript analysis, including surveys and other social-science approaches to measure how AI changes people after they close the app.

read more in theverge.com