CtrlK
The Anthropic Institute is a research organization within Anthropic dedicated to studying the societal challenges posed by increasingly powerful artificial intelligence systems. Announced on March 11, 2026, the Institute consolidates three of Anthropic's existing research groups under co-founder Jack Clark, who assumed a new role as Anthropic's Head of Public Benefit. The Institute operates as an internal think tank with an interdisciplinary staff of machine learning engineers, economists, and social scientists, and its stated mission is to confront the most significant challenges that powerful AI will impose on societies worldwide.
Anthropic was founded in 2021 by Dario Amodei and Daniela Amodei, along with several former OpenAI researchers, as a public benefit corporation focused on AI safety. The company is best known for developing the Claude family of large language models. From its earliest days, Anthropic invested in research that went beyond model capabilities, including work on AI alignment, interpretability, and red-teaming. Over time, three distinct research teams formed around questions about AI's real-world effects: the Frontier Red Team, the Societal Impacts team, and the Economic Research team.
By late 2025 and early 2026, the pace of AI development had accelerated to a degree that made these questions more urgent. AI systems were being adopted across industries at a rapid clip, and policymakers, businesses, and workers were grappling with the implications. According to Jack Clark, he had been considering the idea of a dedicated institute and a shift in his own role since November 2025. The formal announcement came on March 11, 2026, via Anthropic's official blog.
The launch coincided with a turbulent period for the company. Just days earlier, on March 9, 2026, Anthropic had filed a lawsuit against the Trump administration after the Pentagon placed the company on a supply chain risk list, a designation historically reserved for foreign adversaries. The dispute reportedly stemmed from Anthropic's refusal to accept contract terms that would have allowed the use of Claude for autonomous weapons and mass domestic surveillance. Clark told The Verge that the Institute launch had been planned well in advance and was not a response to the Pentagon controversy.
The Anthropic Institute describes its core purpose as telling the world what Anthropic is learning about AI's societal challenges as it builds frontier AI systems, and partnering with external audiences to help address the risks that must be confronted. The Institute draws on its position inside a frontier AI lab, which gives it access to information that only the builders of such systems possess.
The Institute's research agenda is organized around several fundamental questions:
Anthropic frames the Institute as a commitment to candid public reporting about the technology it is building, even when findings are inconvenient. This transparency mandate distinguishes it from internal safety teams at other AI labs, which typically publish their findings selectively. The Institute also pledges to engage directly with groups that expect disruption from AI adoption, including workers and industries facing displacement, and communities anticipating rapid change.
Jack Clark leads the Anthropic Institute in his capacity as Head of Public Benefit. Clark is one of Anthropic's co-founders and has built a career that spans journalism, policy, and AI research.
Before entering the AI industry, Clark worked as a technology journalist. He covered distributed systems as a reporter at The Register and later became what Bloomberg described as the publication's neural network reporter, writing about AI research for Bloomberg BusinessWeek. This journalism background gave him an unusually broad view of the AI landscape and a commitment to making technical developments accessible to non-specialist audiences.
Clark joined OpenAI in its early years, serving as Strategy and Communications Director before becoming Policy Director. In that role he helped shape OpenAI's approach to public communication about AI capabilities and risks. In 2021, he left OpenAI to co-found Anthropic alongside Dario Amodei, Daniela Amodei, and other former OpenAI researchers.
Outside Anthropic, Clark authors Import AI, a weekly newsletter covering AI research that reaches approximately 70,000 subscribers. He was a founding member of the AI Index at Stanford University (2017 to 2024), a project affiliated with Stanford's One Hundred Year Study on AI that tracks progress and trends across the field. He served as an inaugural member of the United States' National Artificial Intelligence Advisory Committee (NAIAC) from 2021 to 2024, and has held advisory roles with the Center for a New American Security (CNAS), the Organization for Economic Co-operation and Development (OECD), and Georgetown University's Center for Security and Emerging Technology (CSET), where he is a non-resident research fellow.
The Anthropic Institute brings together and expands three existing research groups within Anthropic. Each team has a distinct mandate, but their work frequently overlaps and informs one another.
The Frontier Red Team stress-tests AI systems to identify the outermost limits of their current capabilities. Red-teaming is a practice borrowed from cybersecurity and military planning, in which a dedicated group attempts to find vulnerabilities and failure modes in a system before they are discovered by adversaries or surface in real-world use.
At Anthropic, the Frontier Red Team examines AI models for potential risks across domains including cybersecurity, biosecurity, and autonomous behavior. One of the team's most prominent projects was a collaboration with Mozilla in early 2026, in which Claude Opus 4.6 was tasked with independently scanning the Firefox browser's C++ codebase for security vulnerabilities. Over the course of two weeks, the model scanned nearly 6,000 C++ files and submitted 112 unique reports. Mozilla assigned 14 of the discovered vulnerabilities as high-severity, representing nearly a fifth of all high-severity Firefox vulnerabilities remediated in 2025. Most of these issues were fixed in Firefox 148. The team also assessed Claude Opus 4.6's ability to develop exploits for the vulnerabilities it discovered, finding that in most cases the model was unsuccessful, producing a crude exploit in only two out of several hundred attempts that cost approximately $4,000 in API credits.
This type of research serves a dual purpose. It helps Anthropic understand the real-world capabilities of its own models, and it provides concrete evidence for policymakers and security researchers about the current state of AI-enabled cyber operations.
The Societal Impacts team is a technical research group that studies how AI is being used in the real world, investigates which values AI systems should hold, and works to anticipate future uses and risks through experiments, training methods, and evaluations. The team collaborates closely with Anthropic's Policy and Safeguards teams.
Key publications from this team include:
| Publication | Date | Key Findings |
|---|---|---|
| "Values in the Wild" | April 2025 | Analyzed 700,000 real-world interactions and discovered 3,307 distinct values expressed by Claude. Found that Claude adapts its values to context, mirroring users in most cases but resisting when core principles are at stake. Published at COLM 2025. |
| "How AI is Transforming Work at Anthropic" | December 2025 | Surveyed and interviewed Anthropic employees, finding that AI use is radically changing software development workflows. Senior engineers reported concerns about skill atrophy, while junior engineers reported asking senior colleagues fewer questions. |
| "Anthropic Interviewer" | December 2025 | Developed an automated interview tool that conducted large-scale qualitative research with 1,250 professionals about their experience working alongside AI. |
| "What 81,000 People Want from AI" | March 2026 | The largest multilingual qualitative study of its kind, analyzing Claude.ai user perspectives on AI capabilities, hopes, and concerns. |
The team also published early foundational work including "Predictability and Surprise in Large Generative Models" (February 2022), which examined how large models exhibit predictable loss through scaling laws but produce unpredictable emergent capabilities, a finding with direct policy implications.
The Economic Research team tracks AI's impact on employment, productivity, and the broader economy. Its flagship initiative is the Anthropic Economic Index, which analyzes millions of Claude conversations to measure actual AI tool usage patterns globally.
The Economic Index introduces what the team calls "economic primitives": five foundational measurements that track how Claude is used over time. These include task complexity, skill level, purpose (work, education, or personal), AI autonomy, and task success.
Key findings from the Economic Research team include:
| Finding | Detail |
|---|---|
| Augmentation vs. automation | On Claude.ai, AI usage leans toward augmentation (57%) over automation (43%). API usage by businesses shows a higher automation rate of approximately 77%. |
| Directive delegation trend | Directive task delegation (users telling Claude to complete tasks autonomously) rose from 27% to 39% over an eight-month period ending in early 2026. |
| Productivity gains | AI reduces task completion time by approximately 80% on average. If universally adopted over 10 years, current models could increase U.S. labor productivity growth by 1.8% annually. |
| Geographic concentration | AI adoption concentrates in wealthier regions, with significant disparities across geographies and enterprise types. |
| Most-exposed occupations | Computer programmers (75% task coverage), customer service representatives, data entry keyers, and medical record specialists rank among the most exposed occupations. |
In March 2026, the team published "Labor Market Impacts of AI: A New Measure and Early Evidence," authored by Maxim Massenkoff and Peter McCrory. The paper introduced a new metric called "observed exposure," designed to quantify not just which tasks large language models could theoretically speed up, but which are already being automated in practice. The study found that while occupations with high AI exposure showed signs of slower projected growth and reduced hiring for entry-level roles, there was no evidence yet of systematic unemployment increases. However, the researchers identified suggestive evidence that hiring had slowed for workers aged 22 to 25 in high-exposure fields, with job-finding rates dropping by approximately 14% compared to 2022 levels. The paper named the scenario of a "Great Recession for white-collar workers" as a possibility that warranted serious attention.
Alongside the three research teams, the Anthropic Institute announced several high-profile hires at its launch.
Matt Botvinick joined the Institute to lead its work on AI and the rule of law. Botvinick brings an exceptionally diverse academic background. He holds an M.D. from Cornell University (with board certification in Psychiatry), an M.A. in Art History from Columbia University, and a Ph.D. in AI and computational neuroscience from Carnegie Mellon University. He held faculty positions at the University of Pennsylvania and Princeton University, where he served as Professor of Neural Computation, before joining Google DeepMind in 2016. At DeepMind, he rose to Senior Director of Research and authored more than 140 peer-reviewed articles spanning AI, deep learning, reinforcement learning, cognitive science, and computational neuroscience. At the time of his hiring, Botvinick was a Resident Fellow at Yale Law School, where he was also pursuing doctoral studies in law. His appointment signals the Institute's intention to address the complex intersection of AI capabilities and legal frameworks.
Anton Korinek joined the Economic Research team on leave from his position as Professor of Economics at the University of Virginia's Department of Economics and Darden School of Business. Korinek is the Faculty Director of the Economics of Transformative AI (EconTAI) Initiative and has held visiting scholar positions at Harvard University, the World Bank, the International Monetary Fund, the Bank for International Settlements, and numerous central banks. He received his Ph.D. from Columbia University in 2007. His research focuses on how transformative AI systems could reshape economic activity, growth, labor markets, and inequality. In 2025, TIME named him to the TIME100 AI list of the most influential people in artificial intelligence. At the Anthropic Institute, Korinek leads an effort to study how advanced AI could fundamentally alter the structure of economic activity.
Zoe Hitzig joined the Institute to connect the economics research program to model training and development. Hitzig received her Ph.D. in economics from Harvard University in 2023, with research focused on privacy, transparency, and algorithms in markets and communication. Before joining Anthropic, she worked as a Research Scientist at OpenAI from 2024 to 2026, where she studied AI's social and economic impacts. She departed OpenAI after publicly objecting to the company's decision to monetize advertising without adequate guardrails for user data. Beyond her economics work, Hitzig is also a published poet, with work appearing in The New Yorker, The New York Review of Books, and The Paris Review. Her role at the Anthropic Institute bridges the gap between empirical economic research and the technical decisions made during model training.
The Anthropic Institute launch was accompanied by an expansion of Anthropic's Public Policy organization, led by Sarah Heck as Head of Public Policy. Heck previously served as Head of Entrepreneurship at Stripe, where she led Stripe Atlas, the company's platform for global entrepreneurs. Before Stripe, she held roles at the White House National Security Council during the Obama administration, where she directed global entrepreneurship and public diplomacy policy. She also held positions at the U.S. Department of State focused on public diplomacy, countering violent extremism, and technology in diplomacy. Heck holds a degree from Georgetown University's Walsh School of Foreign Service.
Under Heck's leadership, Anthropic plans to triple its policy team and open a permanent office in Washington, D.C. in spring 2026. The policy team includes registered lobbyists from both major U.S. political parties and will continue advocating for export controls on advanced chips, a clear federal AI regulation framework, energy ratepayer protections, and model transparency.
The Anthropic Institute is closely tied to Anthropic's broader identity as a safety-focused AI company. Anthropic was incorporated as a public benefit corporation and established the Long-Term Benefit Trust (LTBT), an independent body of five financially disinterested trustees with authority to select and remove a growing portion of the company's board, eventually comprising a majority. The LTBT trustees have backgrounds in AI safety, national security, public policy, and social enterprise.
Anthropic also maintains its Responsible Scaling Policy (RSP), first published in September 2023 and updated to version 3.0 in February 2026. The RSP defines safety levels for AI models and commits the company to not training or deploying models capable of causing catastrophic harm unless safety and security measures keep risks below acceptable thresholds.
The Anthropic Institute complements these governance mechanisms. While the RSP and LTBT focus on internal decision-making and model deployment, the Institute focuses outward, generating and publishing research that informs external stakeholders. The Institute's access to internal data about model capabilities, user behavior, and economic effects provides it with a vantage point that independent think tanks and academic researchers cannot replicate.
The Anthropic Institute occupies a distinctive position in the landscape of organizations studying AI's societal effects. Several characteristics set it apart.
Embedded within a frontier lab. Unlike independent think tanks such as the Center for AI Safety, the Partnership on AI, or university research centers like Stanford's Human-Centered AI Institute (HAI), the Anthropic Institute sits inside the company that builds frontier AI models. This gives its researchers direct access to model internals, usage data, and capability assessments that external organizations cannot obtain. The tradeoff is that the Institute's independence could be questioned, since it is funded by and reports to the same company whose products it evaluates.
Interdisciplinary by design. Many AI safety teams at other labs focus primarily on technical alignment research. Google DeepMind has its safety division, and OpenAI has maintained various safety-oriented teams (including its former Preparedness team). The Anthropic Institute explicitly combines machine learning engineers with economists and social scientists, reflecting a view that AI's societal effects cannot be understood through a purely technical lens.
Public-facing research mandate. The Institute commits to publishing its findings for external researchers and the public. While other AI labs publish selected research papers, the Anthropic Institute frames public reporting as a central obligation rather than an optional activity.
Engagement with affected communities. The Institute's founding announcement specifically mentions dialogue with workers and industries facing displacement, and with communities anticipating rapid change. This stakeholder engagement model is more commonly associated with policy organizations than with corporate research labs.
| Feature | Anthropic Institute | Independent Think Tanks | Other Lab Safety Teams |
|---|---|---|---|
| Access to frontier model data | Direct internal access | Limited or no access | Internal access |
| Disciplinary scope | ML engineers, economists, social scientists | Varies; often policy-focused | Primarily ML and security |
| Publishing mandate | Central to mission | Central to mission | Selective |
| Funding source | Anthropic | Grants, donations, government | Parent company |
| Community engagement | Explicit commitment | Varies | Rare |
| Independence from lab | Limited; internal unit | Fully independent | Limited; internal unit |
Beyond its established research streams, the Anthropic Institute has signaled several new directions.
Forecasting AI progress. The Institute is developing methods to predict the trajectory of AI capabilities. This work builds on Clark's long involvement with the Stanford AI Index and reflects a growing recognition among policymakers that anticipating the pace of AI development is critical for effective regulation and planning.
AI and the legal system. Under Matt Botvinick's leadership, the Institute plans to study how powerful AI systems could interact with and potentially disrupt legal institutions. This includes questions about AI-generated evidence, automated legal reasoning, liability for AI-caused harms, and the implications of AI systems that can process and generate legal documents at scale.
Recursive self-improvement governance. The Institute's research agenda explicitly includes the question of how to govern AI systems if they develop the ability to improve themselves, a scenario that many AI researchers consider plausible in the near to medium term. This connects to broader debates about AI existential risk and the adequacy of current governance frameworks.
Labor market monitoring. Building on the Economic Research team's work, the Institute plans to continue tracking AI's effects on employment in real time, providing a data-driven counterpoint to speculative predictions about job displacement.
The Anthropic Institute's creation reflects a broader trend in the AI industry toward institutional responses to societal concerns. As AI systems become more capable and more widely deployed, the gap between what these systems can do and what societies are prepared to manage continues to widen. Traditional academic research cycles are too slow to keep pace with the speed of AI development, and independent policy organizations often lack access to the data needed to make informed assessments.
By embedding a research institute inside a frontier AI lab, Anthropic is betting that the most useful societal research about AI will come from the people closest to the technology. This is a significant institutional experiment. If the Institute produces honest, rigorous research that sometimes reflects unfavorably on Anthropic's own products, it could set a new standard for corporate responsibility in the AI industry. If its output is perceived as promotional or self-serving, it could undermine trust in corporate AI safety research more broadly.
The Institute's launch at a moment when Anthropic was locked in a legal battle with the U.S. Department of Defense over AI ethics adds an additional layer of significance. Whether intentional or not, the timing reinforced Anthropic's positioning as a company willing to prioritize safety and societal impact over government contracts and revenue.