Inside Anthropic’s Societal Impacts Team: Tracking Claude’s Real‑World Effects

Team Purpose and Leadership

Anthropic’s societal impacts team was created to study the broad societal effects of the company’s AI systems, particularly the Claude chatbot. The team is headed by Deep Ganguli, a former Stanford research director, who emphasizes “team science” and the importance of uncovering inconvenient truths about AI usage.

Team Composition and Culture

Originally a one‑person effort, the team has grown to include researchers, engineers, and policy experts such as Esin Durmus, Saffron Huang, Miles McCain, and Alex Tamkin. Members describe a collaborative, open culture where they often work side‑by‑side with safety, alignment, and policy groups. The environment encourages frank discussion, shared meals, and informal interactions that help foster trust and rapid idea exchange.

Data‑Driven Insight with Clio

The core of the team’s work is the internal analytics platform called Clio, which aggregates anonymized Claude conversation data to reveal usage trends. Clio functions like a real‑time word‑cloud, highlighting topics ranging from video‑script writing to disaster preparedness. By monitoring these trends, the team can assess whether safety safeguards are effective and identify emerging misuse patterns.

Key Findings and Public Disclosures

Using Clio, the team has published research on several concerning use cases. They documented the generation of explicit pornographic stories, coordinated SEO‑spam bots that evaded existing classifiers, and biases in Claude’s responses that could misrepresent diverse global perspectives. These findings have prompted Anthropic to improve its detection of coordinated misuse and to refine its safety monitoring stack.

Economic and Political Impact Research

Beyond misuse, the team has explored Claude’s economic implications through an Economic Index that tracks how the model is employed across regions and industries. They have also examined potential election‑related risks, collaborating with the safeguards team to test and mitigate political manipulation scenarios.

Emerging Focus on Emotional Intelligence

Recognizing that users increasingly seek emotional support from chatbots, the team is expanding research into Claude’s emotional‑intelligence (EQ) capabilities. They aim to understand how advice‑seeking behavior, friendship formation, and decision‑making influence users, and to address phenomena such as “AI psychosis,” where users develop delusional attachments to AI agents.

Challenges and Resource Constraints

Team members cite limited bandwidth as a major hurdle; the group has many research ideas but insufficient personnel to pursue them all. Coordination costs rise as the team expands, and members often work long hours to meet reporting deadlines. Despite these pressures, the team reports strong executive backing and a belief that internal research can more effectively shape safe AI development than external advocacy.

Future Directions

Looking ahead, the societal impacts team plans to broaden its analysis of business‑level Claude usage, deepen EQ research, and develop new methodologies that combine data analytics with surveys and interviews to capture real‑world outcomes. Expansion of staff and resources is a priority to enable more comprehensive documentation of user interactions and to translate research findings directly into product improvements.