In May 2020, as global lockdowns peaked, Deep Ganguli, then at Stanford’s AI center, felt a growing unease. OpenAI’s unveiling of GPT-3—an AI model exponentially more advanced than anything before—was not just a leap forward; it was a sign of relentless acceleration. Ganguli considered the societal implications, recognizing that the pace of change was outpacing preparedness. He wanted to steer this evolution, not simply observe it.
This led him to Anthropic, a startup founded by ex-OpenAI employees who shared his concerns. Anthropic’s founders believed AI safety wasn’t being prioritized enough, and they wanted to build a company that put humanity first. Jack Clark, a former OpenAI policy director, hired Ganguli to lead a small, critical team: the societal impacts group. Their mission? Ensure AI “interacts positively with people,” from everyday interactions to global politics.
The Core of AI Risk Assessment
Over four years, Ganguli built a team of just nine people within Anthropic’s 2,000-strong workforce. This group tackles the thorniest questions AI poses: its economic consequences, persuasive power, election interference potential, and inherent biases. They’ve published research on these topics, shaping Anthropic’s image as the “safe” AI giant.
However, their influence hinges on a precarious balance. In an industry where speed and profit dominate, a small team dedicated to long-term human impact may not have indefinite freedom. The question is whether executives, ultimately driven by financial goals, will listen to warnings that could slow growth or reveal vulnerabilities.
Transparency as a Strategy
The societal impacts team’s approach is simple: “We are going to tell the truth.” Unlike other AI companies focusing on preventing obvious harms like scams or weaponization, Ganguli’s team seeks “inconvenient truths” that corporations have little incentive to publicize. They share their findings internally and externally, believing transparency builds trust with the public and policymakers.
This includes candidly admitting shortcomings. The team openly discussed how Anthropic’s systems were exploited to create explicit content and spam, publishing the research to help other companies address similar weaknesses. This willingness to expose failures, rather than hiding them, is rare in an industry often defined by secrecy.
Building a Collaborative Culture
The team fosters a collaborative environment, working closely with other departments. Their work is driven by data, but they also emphasize open communication. They even have a Slack channel called Clio Alerts, where automated flags on AI misuse are shared, and members casually collaborate in the office gym or on day trips to the beach.
This camaraderie is intentional. By building deep personal connections, the team creates a space where dissenting opinions are welcomed, and uncomfortable findings are openly discussed. The team also uses a lighthearted “cone of uncertainty” metaphor—reinforced by an actual traffic cone with googly eyes—to acknowledge the limits of their understanding.
Tracking AI’s Real-World Usage: The Clio System
One of the team’s biggest achievements is Clio, Anthropic’s AI usage tracking system. Designed to monitor how people interact with Claude, Anthropic’s chatbot, Clio resembles a real-time word cloud showing trending topics, from writing scripts to solving math problems. The system helps identify both intended and unintended uses, including harmful ones.
Clio’s transparency is deliberate. Anthropic shares the data with internal teams and publishes findings, even if they paint the company in a negative light. This approach has led to improvements in safety monitoring, including better detection of coordinated misuse and proactive identification of emerging risks.
The Challenge of Long-Term Influence
Despite their commitment to transparency, the team operates within Anthropic’s corporate structure. Access to workspaces is restricted, and research findings require approval. This creates a tension between the team’s desire for openness and the company’s need for control.
However, team members remain optimistic. They believe Anthropic’s leadership genuinely values their work, even if direct communication with executives is infrequent. The team’s success depends on maintaining this trust and continuing to deliver data-driven insights that shape Anthropic’s long-term strategy.
Ultimately, the societal impacts team represents a unique experiment: a small group tasked with navigating the human consequences of one of the world’s most disruptive technologies. Their future hinges on whether Anthropic can balance its growth ambitions with its commitment to responsible AI development.
