The Anthropic Institute (TAI) published its research agenda on May 7, 2026, identifying four principal areas of study: economic diffusion, threats and resilience, AI systems in the wild, and AI‑driven R&D. TAI intends to use observations available from a frontier lab to investigate how advanced AI systems change economies, institutions and security, and to share those findings publicly to inform governments, organizations and the general public.
Why this research?
TAI argues that investigating impacts — like studying safety — requires direct contact with frontier models and systems. Anthropic’s internal experience already indicates substantial shifts: software engineering roles are transforming, Anthropic’s internal economy is adjusting, new threats tied to built systems are emerging, and AI appears to be accelerating AI research itself.
Public data and tools
TAI will publish research, data and tools so that external researchers and institutions can work on these questions. Concrete commitments include:
- providing more granular and higher‑cadence outputs from the Anthropic Economic Index about labor impacts and AI usage as an early warning for significant disruption;
- researching which societal areas most need investment to bolster resilience against new AI‑enabled security risks;
- publishing more detailed accounts of how Anthropic’s own R&D has accelerated due to new AI tools and exploring the implications of potential recursive self‑improvement.
Programs and institutional links
The institute’s work is intended to shape Anthropic’s decisions. This may mean publishing data that otherwise would remain internal (e.g., Economic Index outputs) or changing release practices for technology (for example, cyber threat analyses that connect to initiatives such as Project Glasswing). TAI’s work is designed to feed into Anthropic’s Long‑Term Benefit Trust (LTBT), whose mission is to ensure Anthropic optimizes its actions for the long‑term benefit of humanity. The agenda was developed with input from the LTBT and Anthropic staff.
TAI describes the agenda as a living document: questions will be refined as evidence accumulates, and the institute welcomes feedback. It also invites researchers to apply for a funded, four‑month Anthropic Fellowship to work on one or more of the agenda’s topics with TAI mentorship.
The four research pillars
Economic diffusion
TAI wants to understand how deploying increasingly powerful AI systems changes economic outcomes and to build the data and forecasting capacity needed to steer deployment toward public benefit. Key research topics include:
- AI adoption and diffusion: who adopts AI and under what conditions? How do countries, regions and cities capture value, and what roles do policy, business models and open‑weight models play?
- Firm‑level adoption: what drives AI adoption at firms, how does AI change optimal scale and concentration, and what are the consequences for markups and labor share?
- Is AI a general‑purpose technology?: does AI follow historical GPT patterns—rapid commercial adoption where private returns are high but slower uptake where social returns exceed private returns—and can policy change that dynamic?
- Productivity and growth: how will AI affect innovation rates and productivity growth, and what pre‑ or re‑distributive mechanisms might spread gains broadly?
- Transaction costs and markets: how does AI affect exchanges and transaction costs, and when do AI agents negotiating on behalf of people improve outcomes?
- Broad labor market impacts: how will jobs and employment change across sectors and regions? The Anthropic Economic Index Survey will provide monthly signals on how workers perceive AI’s effects on their work, supplemented by higher‑frequency Economic Index data.
- Modulating diffusion: are there industry‑ or government‑level “dials” analogous to monetary policy tools that could regulate the pace of AI diffusion across sectors?
- Future of work and professional pipelines: how will junior roles that historically trained senior experts be affected if AI automates formative tasks, and what will people need to study to remain well‑positioned?
Threats and resilience
TAI emphasizes that AI systems are often dual‑use: capability improvements that enable benefits may also enable misuse (e.g., in biology or cyber). Research aims to strengthen global resilience and build early warning systems for emergent threats. Focus areas include:
- Assessing dual‑use risks and observability: can we develop tools to detect whether AI is being used for harmful ends?
- Pricing risk: what market‑driven or technical approaches can increase societal resilience to anticipated AI threats?
- Offense‑defense balance: will AI structurally favor attackers in domains like cyber and bio, and how will AI change the nature of conflict?
- Crisis planning and geopolitical infrastructure: what mechanisms would be required for crisis coordination involving AI systems, which might be company‑to‑company or company‑to‑state rather than purely state‑to‑state?
- Faster defensive responses: given that AI capabilities can advance in months while regulatory and infrastructure responses take years, how can defenses—such as automated patching, AI‑enabled threat detection, or pre‑positioned responses—keep pace?
- Surveillance capabilities: how will AI change the cost and effectiveness of surveillance?
Many of these questions will guide the work of TAI’s Frontier Red Team.
AI systems in the wild
This pillar studies how people and institutions actually interact with deployed AI systems. The Societal Impacts team will develop observational software and large‑scale qualitative research tools to study these dynamics. Topics include:
- Group epistemology: what happens to collective beliefs, writing styles and problem‑solving when many people consult the same models?
- Critical thinking: how to detect and prevent erosion of human critical reasoning as AI systems become more trusted?
- Technological interfaces: which interfaces promote human agency versus passivity?
- Managing human–AI systems: how can mixed teams of humans and AIs be managed, and how might AIs manage teams?
- Behavioral effects and measurement: what monitoring can identify AI‑driven behavioral shifts at scale?
- Enabling research via transparency: which transparency regimes and tools allow a broader set of researchers to study real‑world AI usage?
- Model values and governance of autonomous agents: how do an AI’s expressed “values” relate to training processes, and which aspects of existing laws and accountability frameworks apply to autonomous agents?
- AI governance of AI: where can AI usefully govern other AI systems, and where must humans remain in the loop?
AI‑driven R&D
As AI systems perform more research tasks autonomously or semi‑autonomously, and potentially help develop successor systems, TAI raises questions about governance and early warning. Research priorities include:
- Governance of AI R&D: how can humans retain meaningful visibility and control over systems that autonomously improve themselves, and what will govern such systems?
- Fire drills for a potential intelligence explosion: what tabletop or exercise scenarios could realistically test decision‑making of lab leadership, boards and governments?
- Telemetry for AI R&D: how can we measure aggregate R&D speed, and what telemetry is needed to detect signs of recursive self‑improvement?
- Controlling AI acceleration: if an intelligence explosion were possible, what intervention points might slow or alter its pace, and which actors should have authority?
- AI for R&D in other fields: how unevenly is AI accelerating different sciences, and what are the implications for which human problems receive attention and resources?
Closing notes
TAI’s agenda builds on Anthropic’s frontier‑lab position and commits to public data sharing. The agenda is explicitly described as evolving, with the institute expecting new questions to appear and inviting external feedback. A funded, four‑month Anthropic Fellowship is available for researchers who want to work on these issues.
Last updated: May 7, 2026.
