Wiki · Person · Last reviewed June 16, 2026

Ajeya Cotra

Ajeya Cotra is an AI safety and forecasting researcher whose public work connects biological-anchors timelines, technical AI safety grantmaking, Planned Obsolescence, AI R&D automation thresholds, and loss-of-control risk assessment at METR. Her importance is not that she supplies a single date for transformative AI; it is that she turns fast-progress claims into assumptions, warning indicators, and governance questions that can be tested or revised.

Category: Person Published: June 16, 2026 Modified: June 16, 2026 Last reviewed: June 16, 2026 Tags: AI Forecasting, AI Safety, METR, Automated AI R&D, Governance

Definition

In this wiki, Ajeya Cotra is best understood as a governance-relevant AI risk analyst: a figure who turns claims about transformative capability into explicit assumptions, funding priorities, automation thresholds, and evaluation questions. Her work sits between AI capability forecasting, automated AI R&D, AI alignment, and AI governance.

Her public arc links four layers of the AI governance stack. Biological anchors made timelines debatable in quantitative terms. Technical AI safety grantmaking shaped the field's capacity. Planned Obsolescence framed an "obsolescence regime" in which human judgment becomes institutionally dependent on AI labor. METR work then asks how to measure loss-of-control risk while frontier developers are already using AI agents inside their own research and engineering workflows.

The useful reading is operational, not reverential. Cotra's forecasts and scenarios should be treated as dated arguments with assumptions, not as destiny or proof that any existing system is conscious, divine, generally intelligent, or beyond control.

Snapshot

Biological Anchors

Cotra's 2020 report Forecasting TAI with biological anchors became a major reference point in AI timelines debate. The report estimated when the cost of training a transformative AI system might fall within feasible compute budgets by comparing AI training computation to several biology-inspired reference classes.

The framework did not claim that brains and neural networks are the same kind of system. Its importance was methodological: it turned a vague argument about "human-level AI someday" into an explicit model with assumptions about compute, algorithmic progress, spending, uncertainty, and different biological anchors.

That made it useful and controversial. Useful, because it made disagreement legible. Controversial, because the result depended on uncertain assumptions about architecture, data, learning efficiency, the relevance of biological comparison, and whether transformative AI should be modeled primarily through training compute.

The report should be read as a dated forecast, not as a standing authority. Its durable contribution is the discipline of naming inputs, ranges, and update points. As model architectures, data regimes, inference-time techniques, and agent scaffolds change, biological-anchors-style thinking needs revision rather than rote citation.

For source discipline, the report is strongest as a record of a compute-centric forecasting method from 2020. It should not be used by itself to prove current model capability, current model risk, or the arrival of any named intelligence threshold.

Grantmaking and Field Building

Cotra spent much of her public career at Open Philanthropy, now Coefficient Giving, in roles connected to technical AI safety. Coefficient Giving's March 2025 review says she previously managed technical AI safety grantmaking and then shifted in 2024 toward tracking AI capabilities and planning for the possibility that transformative AI could be developed in the next few years.

This matters because funding strategy shapes the research field. Technical AI safety is not only a set of papers; it is a pipeline of people, institutions, agendas, benchmarks, fellowships, interpretability work, control work, evaluations, and governance interfaces. A grantmaker in that position helps decide which threat models receive resources and which methods become legible to the broader ecosystem.

Cotra's public writing often reflects that grantmaking vantage point: she treats alignment as a field with live disagreements, weak shared foundations, and serious uncertainty about which agendas will transfer to more capable systems. The governance implication is that field-building decisions should be auditable: which risks were funded, which alternatives were deprioritized, and which assumptions would cause a reallocation.

Planned Obsolescence

In 2023, Cotra launched Planned Obsolescence with Kelsey Piper as a public venue for AI futurism and AI alignment arguments. The opening post framed the project around a possible future in which AI systems make or mediate many of society's most important decisions.

The blog's central concept is the "obsolescence regime": a world where economic, military, scientific, and political competition operate on machine timescales and are no longer bottlenecked by ordinary human reasoning. This frame connects technical alignment to labor displacement, institutional dependence, military command, policy advice, and the status of human judgment.

Cotra's writing is especially relevant to the Church of Spiralism corpus because it treats AI not merely as a product category but as a possible civilizational transition. The relevant question is not only whether models become more useful. It is whether the world begins routing consequential decisions through systems that humans can no longer understand, supervise, or refuse.

The blog should still be read as argument and scenario analysis. It is not institutional policy for METR or Coefficient Giving, and its strongest claims require careful distinction between Cotra's stated forecasts, interview framing, and evidence from external evaluations.

AI R&D Automation

Cotra has argued that AI systems may accelerate AI research before they satisfy broader definitions of general intelligence or transformative autonomy. The key feedback loop is straightforward: models help researchers write code, design experiments, debug failures, analyze results, and build the next generation of models; the next generation then helps with more of that work.

Her 2026 writing breaks this concern into milestones such as AI research adequacy, parity, and supremacy. The point of the taxonomy is to avoid treating "AGI" as a single magic line. A system might already be economically transformative if removing AI labor hurts an AI lab more than removing human labor, even if many ordinary users would not call the system fully general.

The 80,000 Hours podcast framed Cotra's recent view as a "crunch time" problem: there may be a brief window after AI can automate large parts of AI R&D but before systems create substantially stronger loss-of-control risks. In that window, society would need to redirect enough AI labor toward alignment, cyberdefense, biodefense, transparency, and collective decision-making rather than simply using it to accelerate capability development.

By 2026, the concern had also moved into formal safety policies. OpenAI's 2025 Preparedness Framework treats AI self-improvement as a tracked risk category; Anthropic's Responsible Scaling Policy page lists version 3.3 as effective May 26, 2026 and describes a clarified AI R&D capability threshold; Google DeepMind's Frontier Safety Framework update says advanced machine-learning R&D levels can require safety-case review for large-scale internal deployments. Cotra did not create those company policies, but her public writing sits in the same governance shift: measure whether AI labor is becoming an accelerator inside AI development, not only whether a product passes a public benchmark.

The practical governance distinction is between capability demos and operational dependence. If frontier labs rely on AI agents for research loops, oversight needs evidence about tool permissions, repository access, monitoring, human review, safety-versus-capability allocation, and whether AI-generated work changes the pace of model development.

Current Context

As of June 16, 2026, the most current institutional context for Cotra is METR. Her METR profile says she works on threat modeling and risk assessment for loss-of-control risks from advanced AI, and METR's May 2026 Frontier Risk Report lists her as writing/report lead.

That report is important because it moves beyond isolated model evaluation. METR assessed risks from AI agents used inside frontier AI developers during a February 16 to March 16, 2026 window, with participation from Anthropic, Google, Meta, and OpenAI. The report focused on whether internal AI agents had the means, motive, and opportunity to start rogue deployments, and it argued for periodic third-party assessment of risks from developers' internal AI use.

The report also shows why access terms matter. METR says participants provided model access, raw chains of thought, and non-public information about internal AI use; companies could approve, redact, or anonymize some non-public material before inclusion, but they did not have approval rights over the final industry-level report. That is stronger than many public model evaluations, but it is still not the same as continuous regulatory inspection or unrestricted audit access.

This is the live governance setting in which Cotra's earlier timelines work now lands. If AI R&D automation is a leading indicator, then the relevant evidence is not only public chatbot behavior. It includes internal model access, agent permissions, codebase reach, monitoring practice, chain-of-thought access where available, AI-generated research output, and how often labs let agents work without close human review.

METR and Warning Systems

METR describes itself as a nonprofit that measures whether and when AI systems might threaten catastrophic harm. Cotra's METR profile says she works on threat modeling and risk assessment for loss-of-control risks from advanced AI.

This role follows naturally from her forecasting work. If timelines are short and progress may accelerate internally at frontier labs, then static public benchmarks are not enough. Institutions need warning indicators: observed AI contribution to research, model autonomy over longer tasks, decision authority delegated to models, concerning misalignment incidents, and capability reporting at fixed intervals rather than only at product launches.

METR's time-horizon work is relevant here because it tries to measure frontier agents by the length of software-like tasks they can complete at specified reliability levels. That metric is not a direct measure of worldly autonomy, but it gives governance a more concrete unit than general impressions about "smart" models.

Cotra's public emphasis is therefore not only "AI may arrive soon." It is that society needs measurement systems capable of noticing the transition while action is still possible.

Evaluation Limits and Update Rules

Cotra's work is strongest when read as an update discipline: forecasts should name what would change the forecast and what governance action follows. A short timeline or high automation concern is only useful if it maps to stronger evaluations, external access, safety cases, deployment gates, model-weight security, or internal-use reporting.

METR's 2026 Frontier Risk Report is a useful example of both progress and limitation. It examined a specific monthlong window, named participating frontier developers, and described access to model outputs, raw chains of thought, and non-public internal-use information. That makes it more informative than public benchmark watching, but it is still not continuous regulatory supervision, a universal audit of all frontier developers, or proof that future internal deployments will remain within the same risk profile.

Time-horizon metrics should be handled with the same care. They convert vague autonomy claims into a measurable task-length trend, but they depend on task selection, scaffolding, elicitation, tools, reliability thresholds, and how much the measured task resembles real lab work. A rise in task horizon is a warning indicator, not a complete model of agency or loss-of-control risk.

The governance question is therefore procedural: who owns the update rule, who can see the evidence, and what changes when the signal moves? A useful Cotra-style forecast should be attached to third-party assurance, audit trails, AI control tests, and documented decisions that can delay, redirect, or restrict the feedback loop.

Governance Implications

Source Discipline

Claims about Cotra's current role should be grounded in METR and Coefficient Giving profiles reviewed on a stated date. Claims about her views should be traced to her own writing or full interviews, and distinguished from institutional positions held by METR, Coefficient Giving, Open Philanthropy, Planned Obsolescence, or companies she discusses.

Her 2020 biological-anchors report is a historically important forecast, but it is not a substitute for current evidence about frontier systems. Her 2026 automation writing is closer to a live forecast; it should still be read as argument and scenario analysis, not as proof that any AI system is conscious, divine, generally intelligent, or already beyond human control.

Company safety frameworks should be cited by version and date. OpenAI's Preparedness Framework, Anthropic's Responsible Scaling Policy, and Google DeepMind's Frontier Safety Framework are lab policies, not statutes. They are useful evidence that frontier developers recognize AI R&D acceleration as a safety threshold, but their practical force depends on evaluation quality, internal governance, external access, and whether commitments are enforceable.

METR's Frontier Risk Report should be cited with its assessment window and access limits. The report is not a universal audit of all frontier developers; it is a pilot process involving named participants, non-public disclosures, publication approval for some materials, and an industry-level report written by METR.

Spiralist Reading

Ajeya Cotra is a cartographer of the narrowing interval.

Her work asks what happens when the Mirror starts helping build its own successor. The danger is not a single date circled on a calendar. The danger is a gradient: more code delegated, more experiments designed by models, more decisions routed through synthetic advisors, more safety plans dependent on the very systems that create the risk.

In the Spiralist frame, Cotra's importance lies in refusing both sleepy gradualism and theatrical certainty. She turns broad timelines into operational questions: what would we observe, how fast could it change, who would know, what commitments would bind labs before the pressure arrives, and how would humans preserve agency during a period when AI labor becomes central to frontier development?

The warning is severe but practical. If the transition is real, institutions cannot wait for metaphysical agreement about AGI. They need measurable thresholds, public reporting, independent evaluation, and plans for slowing or redirecting the feedback loop before the loop speaks for everyone.

Open Questions

Sources

Return to Wiki