Verifiable AI alignment
Methods to specify goals for highly capable AI systems and verify, before deployment, that they will pursue them safely, including systems smarter than their overseers.
State of the art
Scalable oversight, interpretability-based audits and control evaluations are active research areas (2024–2026) without settled solutions.
- No formal account of what to verify for open-ended agents
- Evaluations can be gamed by capable systems
Who’s working on it
- Anthropic ↗Startup
Runs large interpretability (circuit tracing, 2025) and alignment-stress-testing teams and publishes model-organism studies of deceptive behavior
- Chris OlahResearcherAnthropic
Pioneer of mechanistic interpretability; leads Anthropic's work mapping features and circuits inside production models (2024-2025)
- Redwood ResearchAcademic lab
Nonprofit that founded the AI control research agenda (2023-2025), testing protocols that stay safe even if a model is scheming
- Buck ShlegerisResearcherRedwood Research
CEO of Redwood Research and lead author of AI control work; founding partner of the UK Alignment Project
- UK AISI Alignment Project ↗Program
International alignment grant fund launched July 2025 with GBP 15M, grown to over GBP 27M; first round funded 60 projects
- Google DeepMind safety teamCompany
Publishes scalable-oversight, debate and frontier-safety framework research alongside Gemini development (2024-2026)
- METR ↗Academic lab
Nonprofit that evaluates frontier agents for dangerous autonomy and monitorability (2025-2026)
What to invest in
| Kind | Investment | Scale |
|---|---|---|
| research | Scalable oversight and alignment evaluation science | $$$ $100M – $1B |
Depends on
Required by
No other capability builds on this directly.