The Anthropic Institute (TAI) is Anthropic's unit for studying the effects of AI on the world from inside a frontier lab and publishing what it learns. This post, published on 7 May 2026, sets out its research agenda as a list of questions. It reports no results. Its premise, taken from Anthropic's earlier "Core Views on AI Safety", is that just as safety research needs close contact with frontier systems, so does research on AI's effects on security, the economy and society. The post states that Anthropic already sees "early evidence that jobs like software engineering are changing radically", shifts in its own internal economy, new threats from the systems it builds, and early signs of AI speeding up AI research itself.
What the Institute says it will do
The post commits TAI to studying real-world impacts and publishing the findings so that outside organizations, governments and the public can make better decisions about AI. It names three kinds of material it will share: more granular and more frequent data from the Anthropic Economic Index, meant to serve as "an early warning signal" for disruption; research on which areas of society most need investment in resilience against AI-enabled security risks; and more detail on how Anthropic's own work has sped up with AI tools, with ideas about the implications of recursive self-improvement.
The post also says TAI will shape Anthropic's own decisions, for example by leading the company to share data it otherwise would not, or to release technology differently, citing cyber threat analyses that feed into Project Glasswing. Its work is expected to inform Anthropic's Long-Term Benefit Trust, with which the agenda was developed. The agenda is described as "a living agenda, rather than a fixed one", last updated on 7 May 2026, and the post invites feedback and applications to a four-month Anthropic Fellowship to work on the questions.
The four research areas
Economic diffusion
This area builds on the Economic Index and asks how deploying more capable AI changes the economy. Its questions fall under four headings:
- Adoption and diffusion. What determines whether a country, region or city can access AI and capture value from it; what drives adoption by firms and how concentrated it is; whether AI follows the pattern of earlier general-purpose technologies, adopted fastest in high-margin commercial uses and slowest where social returns exceed private returns. One question asks what happens to industrial organization if "a 3-person team or company can now do what required 300 before".
- Productivity and growth. AI's effect on innovation and productivity; mechanisms to share the gains; how AI agents negotiating on people's behalf affect market efficiency and fairness.
- Labor markets. How jobs and employment will change and vary by region; whether AI companies and governments could have "dials" to modulate the rate of diffusion sector by sector, by analogy with central bank policy rates. The Anthropic Economic Index Survey is named as a source of monthly signals here.
- Jobs and workplaces. How workers experience changes in their professions and how much influence they have; how people become experts if AI absorbs the junior tasks that used to train them; what people should study; and what would let people find meaning if paid work becomes less central.
Threats and resilience
This area starts from the observation that AI systems advance many capabilities at once, including dual-use ones: a system better at biology is better at bioweapons, and one better at programming is better at hacking. Many of its questions are said to drive the agenda of Anthropic's Frontier Red Team. They cover observability tools for surveillance and repression, market-based ways to price AI risk, whether AI structurally favors attackers in cyber, bio or military command and control, what crisis infrastructure (by analogy with the Cold War hotline) an AI crisis would need, how defenses that move on a timescale of years can keep up with capabilities that advance in months, and how AI changes surveillance.
AI systems in the wild
This area, linked to Anthropic's Societal Impacts team, studies how people and institutions change through interacting with AI. Its questions concern group epistemology when many people consult the same few models, the risk that deference to AI degrades critical thinking, interfaces that promote human agency, managing mixed teams of humans and AI (and AI managing humans), behavioral effects comparable to those of social media, and transparency regimes that would let researchers outside AI companies study real usage.
A subsection on understanding and governing models includes the question most directly about written principles: "how can we measure the influence that an AI 'constitution' has on behavior of the model once deployed?" The post says TAI will extend its previous research on expressed values, linking Values in the Wild. Other questions in the subsection ask which parts of existing law could apply to autonomous agents (its example is how naval law treats abandoned ships), whether agents could reliably carry a unique identity, where humans should stay "in the loop" when AI is used to govern AI, and what norms emerge when agents interact.
AI-driven R&D
The last area concerns AI systems that do research, including research on AI itself, which the post calls "AI-driven AI R&D". It suggests this may be a "natural dividend" of more capable systems, in the way coding ability brought dual-use cyber ability, and says it "holds within itself the potential for significant danger". Its questions ask how humans keep visibility into and control over self-improving systems, how to run a "fire drill" or tabletop exercise for an intelligence explosion that tests lab leadership, boards and governments, what telemetry would measure the aggregate speed of AI R&D and warn of recursive self-improvement, and where intervention points would lie if acceleration needed to be slowed, and who should hold them. A second group asks how unevenly AI is speeding up different sciences and whether model improvement can be steered toward domains with large social benefit, such as drug discovery and materials science, that markets underinvest in.
Limitations
The post is an agenda, not a report. It lists questions without methods, timelines, or criteria for when a question counts as answered, and it says the list will change as evidence accumulates. The commitments to publish are stated intentions. The post does not say how TAI's findings will be weighed in Anthropic's decisions beyond the examples given.
Research questions
These ask what evidence would show how the agenda is carried out; the agenda's own questions are listed above.
- Which agenda questions lead to published work, and how quickly? Tracking TAI publications against the May 2026 list, and noting revisions to the list, would show which areas receive attention.
- How would the influence of a constitution on deployed behavior be measured? The agenda poses the question without a method. Comparing measured values in real conversations before and after a published change to the constitution, alongside training-data audits, is one design the question invites.
- Does higher-frequency Economic Index data give earlier warning than official statistics? Comparing the timing of changes in TAI's releases with later movements in government labor data would test the "early warning" aim.
- What would AI R&D telemetry look like in practice? Proposals for measuring the speed of AI research inside a lab, and their validation against known speedups, would show whether the agenda's telemetry question can be answered with available data.
How it connects
- Announcing the Anthropic Economic Index Survey (April 2026) is cited in the agenda as a monthly source for its questions on AI and jobs.
- Values in the Wild (2025) is the earlier research on expressed values that the agenda says it will extend.
References
- Anthropic (2026). Focus areas for The Anthropic Institute. Policy, 7 May 2026. https://www.anthropic.com/research/anthropic-institute-agenda