Anthropic Institute Research Agenda: Four Focus Areas for Studying AI's Societal Impact
The Anthropic Institute (TAI) has published the research agenda it will pursue from inside a frontier AI lab, organized around four focus areas: economic diffusion, threats and resilience, AI systems in the wild, and AI-driven R&D. The agenda matters because it commits Anthropic to publishing labor-market signals, resilience research, and internal evidence about how AI is accelerating its own R&D — data the company says will increasingly feed into both its own decisions and its Long-Term Benefit Trust.
What The Anthropic Institute is and why it exists
The Anthropic Institute is an internal Anthropic organization that studies AI's real-world impacts using information available from within a frontier lab, then publishes its findings for outside organizations, governments, and the public. The stated rationale is methodological: Anthropic argues in Core Views on AI Safety that effective safety research requires close contact with frontier AI systems, and TAI applies the same logic to research on AI's effects on security, the economy, and society.
Anthropic states it already observes early signals it wants to study: jobs such as software engineering are changing radically, the internal economy of Anthropic is starting to shift, new threats are emerging from the systems the company builds, and there are early signs of AI contributing to speeding up AI research and development itself. TAI's purpose is to research how these dynamics might shape the outside world and how the public can help direct those changes.
The four research pillars at a glance
TAI's agenda is divided into four areas:
- Economic diffusion: how AI deployment changes economies, labor markets, and the future of work.
- Threats and resilience: how AI changes the threat landscape, including dual-use capabilities, and how societies can prepare.
- AI systems in the wild: how people and institutions change when they interact with AI systems.
- AI-driven R&D: how AI accelerates scientific research and the development of AI itself.
The document states that the agenda is a living one, that it was developed together with Anthropic's Long-Term Benefit Trust and staff across the company, and that it will be revised as evidence accumulates. The page carries a last-updated stamp of May 7, 2026.
Economic diffusion: measuring how AI changes work
This pillar asks how increasingly powerful AI systems change the economy and how to build the economic data and predictive ability needed to deploy AI in ways that benefit the public. TAI says it will further develop the data in The Anthropic Economic Index and explore other methods for modeling effects such as job loss or unprecedented economic growth.
The pillar is split into four question clusters:
AI adoption and diffusion
Questions include who adopts AI and what determines whether a country, region, or city can access it and capture economic value from it; how free or open-weight models change that dynamic; what causes firm-level adoption and how concentrated AI usage is across firms; whether a three-person team or company doing what previously required 300 changes industrial organization; and whether large-scale centralization of knowledge gives firms stronger incentives to systematically surveil workers. TAI also asks whether AI follows the pattern of previous general purpose technologies, where adoption is fastest in high-margin commercial applications and slowest where social returns exceed private returns.
Productivity and economic growth
TAI asks what impact AI will have on the rate of innovation and productivity growth across the economy, what pre- or re-distributive mechanisms could spread gains from AI more broadly, and how AI changes transaction costs and systems of exchange in marketplaces — including when agents that negotiate on a person's behalf improve market efficiency and equitable outcomes, and when they do not.
Broad labor market impacts
TAI asks how AI will change jobs and employment across different parts of the economy, what new tasks and jobs could emerge, and how changes will vary across regions and countries. The Anthropic Economic Index Survey is described as a monthly signal of how people see AI affecting their work and what they expect for the future, and TAI says it is updating the Economic Index to share more high-frequency, granular data. A further question asks whether AI diffusion can be modulated the way central banks moderate inflation through dials such as the policy rate and forward guidance — that is, whether AI companies, at an industry level and in partnership with government, might turn analogous dials to control the rate of diffusion sector by sector.
The future of jobs and workplaces
This cluster covers how workers experience changes in their professions and whether worker power can be preserved or transformed; how professions that rely on junior roles such as paralegals, junior analysts, and associate developers will train senior practitioners of the future if AI absorbs the tasks that historically built expertise; what people should study to be well positioned for the future; and, if AI substantially reduces the centrality of paid work, what conditions allow people to reallocate time toward other sources of meaning.
Threats and resilience: dual-use capabilities and the defense gap
This pillar starts from the observation that AI systems advance many capabilities at once, including dual-use capabilities: a system that improves at biology also improves at creating biological weapons, and a system good at programming is also better at hacking computers. TAI's goals are to build partnerships that improve societal resilience to transformative AI and to develop early warning systems for new threats. Many of these questions will drive the research agenda of Anthropic's Frontier Red Team.
Assessing risk and dual-use capabilities
Questions include whether observability tools can show if and how dual-use tools enable surveillance and repression; whether market-driven approaches such as new ways of pricing risk can improve resilience ahead of predictable threats like improved AI cyberattack capabilities; and whether AI-enabled capabilities structurally favor attackers in domains such as cyber and bio, including when AI is integrated into command and control systems.
Establishing risk mitigations
TAI asks what geopolitical infrastructure would be needed in an AI crisis, drawing the analogy of the Cold War hotline between the American president and the Kremlin and noting such infrastructure might be company-to-state or company-to-company rather than state-to-state. A second question concerns tempo: AI capabilities can advance in months, while regulatory, insurance, and infrastructure responses operate on timescales of years. TAI asks how that gap can be closed, whether defensive mechanisms such as automated patching, AI-enabled threat detection, or pre-positioned response capabilities can match the tempo and scale of AI-enabled offense, or whether the asymmetry is structural.
Intelligence capabilities for surveillance
A smaller cluster asks how AI changes how surveillance works — whether it makes surveillance cheaper, more effective, or both.
AI systems in the wild: how people and institutions change
This pillar, described as a core focus area for Anthropic's Societal Impacts team, examines the interaction between people and organizations and AI systems. TAI says it is advancing existing tools and building new ones for this work, ranging from software for better observability of Anthropic's platform to tools for conducting large-scale qualitative surveys.
Impact on individuals and societies
TAI asks what happens to epistemology when a large fraction of a population consults the same few models, and whether large-scale changes in beliefs, writing style, and problem-solving approaches attributable to shared AI use can be measured. It asks how to detect and avoid degradation of human critical thinking as people defer more to AI judgment. It asks what interfaces can be built so that AI systems promote rather than erode human agency, using the contrast between passive television viewing and generative computer use. It also asks how humans can manage teams mixing humans and AI systems, and how that relationship could be inverted.
Identifying significant impacts
Questions include what monitoring or measurement can inform researchers about AI shaping human behavior, analogous to behavioral changes attributed to social media, and whether transparency regimes and tools can let people beyond frontier AI companies study real-world AI usage.
Understanding and governing AI models
TAI asks what the expressed values of AI systems are, how they relate to how the systems were trained, and how to measure the influence of an AI constitution on deployed behavior — extending previous work on values in the wild. On governance, it asks which aspects of existing laws, governance systems, and accountability mechanisms could be adapted to autonomous agents, citing the example that naval law's treatment of abandoned ships is relevant to agents running without human oversight, and conversely which existing law already applies and should not be changed. On reliability, it asks whether AI agents can be given a unique identity they reliably output even without direct human control. It also asks how effectively AI can be used to govern AI systems and where humans have a comparative advantage or a legal or normative requirement to remain in the loop, plus what norms emerge when AI agents interact with one another.
AI-driven R&D: measuring the speed of self-improvement
This pillar concerns research carried out autonomously or semi-autonomously by AI systems with less active human oversight — including AI research in which increasingly powerful systems help develop successor versions of themselves, which Anthropic calls AI-driven AI R&D. Anthropic describes this as potentially a natural dividend of making more capable systems, comparable to how coding advances yield dual-use cyber capabilities and scientific advances may yield dual-use bio capabilities, while also stating that AI-driven AI R&D holds within itself the potential for significant danger.
AI for AI R&D
TAI asks how humans exercise meaningful visibility into and control over systems that autonomously develop and improve themselves, and what will eventually govern them. It asks what a fire drill for an intelligence explosion would look like as a tabletop exercise testing the decision-making of lab leadership, boards, and governments. It asks how to measure the aggregate speed of AI R&D, what telemetry and technical affordances are required, and whether metrics of AI R&D could serve as early warning signals for recursive self-improvement. It also asks what intervention points could slow or change the rate of an intelligence explosion, and which entities — governments or companies — should wield that capacity.
AI for R&D in general
On AI-driven research in other fields, TAI asks how uneven the acceleration of science is, noting that it depends on data availability, evaluation signals, and how much knowledge is tacit or institutionally gated, and what the changing composition of scientific progress implies for which human problems get solved first. It also asks whether models can be improved to address social externalities, observing that domains with large positive externalities such as drug discovery and materials science receive less investment than their value warrants because markets steer model improvement according to private return.
What TAI commits to publishing
Beyond posing questions, the agenda names three concrete categories of output TAI intends to share:
- More granular information from The Anthropic Economic Index, at a higher cadence, about labor impacts and AI usage, intended as an early warning signal for significant change and disruption.
- Research on the societal areas most in need of resilience investment in the face of new AI-enabled security risks.
- More detailed information about how work at Anthropic has sped up because of new AI tools, and ideas about the implications of potential recursive self-improvement of AI systems.
How TAI connects to Anthropic's governance
The agenda states that TAI will shape decisions Anthropic makes. The examples given are the company sharing data it otherwise would not, such as the Economic Index, and approaching how it releases technology differently, such as cyber threat analyses that feed into initiatives like Project Glasswing. The document also states that work developed by TAI is expected to increasingly serve as important inputs to Anthropic's Long-Term Benefit Trust, whose mission is to ensure Anthropic continually optimizes its actions for the long-term benefit of humanity. The agenda was developed with the LTBT and with staff across Anthropic.
A living agenda and the Anthropic Fellowship
Anthropic describes this as a living agenda rather than a fixed one, writing: "We'll continue to fine-tune these questions as evidence accumulates, and we expect new questions to emerge that aren't captured here today." The company welcomes feedback and says it will revise the agenda in light of what it learns.
For outside researchers who want to work on these questions directly, Anthropic points to its Anthropic Fellow program: a four-month funded opportunity to tackle one or more of the agenda's questions with mentorship from TAI team members, with applications handled through Anthropic's job board.
What the document does not contain
The published agenda is a set of research questions and publication commitments. It contains no reported results, benchmarks, timelines, or quantitative targets for any of the four pillars, and it does not specify which questions will be answered first. The concrete artifacts named in the post are the Anthropic Economic Index, the Anthropic Economic Index Survey, the Frontier Red Team, the Societal Impacts team, Project Glasswing, and the Anthropic Fellowship.