Who we’re looking for
Experts in one or more of these AI capabilities:
- Autonomous AI R&D and recursive self-improvement
- Autonomous replication and adaptation
- Situational awareness
- Deceptive misalignment
- Undermining of safeguards
To take part, you need both:
- At least three years in research, evaluation, policy or strategy roles in AI safety, evaluation or capabilities research relevant to these capabilities. Research fellowships and a relevant PhD count.
- At least three public contributions from that work, such as papers, preprints, reports, model cards or safety frameworks, with at least one on one of these capabilities.
What taking part involves
Three online surveys, about 3 weeks apart (~Nov 2026 - Jan 2027), each taking about 30 minutes. In them, you will:
- Judge whether each domain needs a red line, and how well each proposed measure could track the capability
- Rate where each threshold should sit, from much too risky to much too conservative, and flag any you think have already been crossed
- Propose your own red lines
- Optionally, comment on other red line domains not considered in this study, the role of propensity in defining red lines, the evidence needed to conclude a red line has been crossed, and minimum conditions for evaluating capabilities
- In later rounds, review anonymised panel results and reasons, revise your ratings if you wish, and rate new thresholds proposed by other experts
How the study works
Capability
Autonomous AI R&D
Measure
KernelBench (Stanford)
Threshold
Writes GPU code faster than standard PyTorch
Does this capability need a red line?
YesNo
How good is this measure?
BadWeakGoodBest
Is this capability too risky, about right or too conservative for a red line?
You
too riskytoo conservative
Rounds 2 and 3
You see how other experts answered and their reasoning, then update your answers
You
too riskytoo conservative
Report
Expert consensus or disagreement, made public to inform decisions
Experts reached consensus that Autonomous AI R&D needs a red line.
Expert consensus is that KernelBench is a good measure of Autonomous AI R&D.
Expert consensus is that this threshold is about right for a red line.
KernelBench is used as an example - not real data
In response to the Global Call for AI Red Lines and calls to slow rapid AI progress, we are conducting a Delphi process with a carefully curated panel of experts in AI capabilities from government agencies (e.g. international AI Safety Institutes), frontier AI companies (e.g. Anthropic, OpenAI), independent research organisations (e.g. METR, Apollo) and other relevant institutions (e.g. universities, evaluators).
By an AI capability red line, we mean capabilities so dangerous that reaching them should trigger a pause or slowdown in frontier AI development.