Interpretability: reading what a model is doing internally What interpretability research actually does, why Anthropic’s CEO called it urgent, and how close the field is to succeeding at reading a model's mind.
Source-checked 5 min
04 Off-ramps
The off-ramps that actually exist
Technical and institutional controls shape whether the severe scenarios in chapter 02 become more or less likely; household preparation cannot substitute for them.
Technical safety levers
- Interpretability research. Inspecting model representations and internal computation as one complement to behavioral evaluations, access controls and deployment monitoring.
- Scalable oversight. Techniques — debate, recursive reward modeling, weak-to-strong generalization — intended to help less capable overseers supervise more capable systems.
- Corrigibility research. Studying how systems might remain receptive to legitimate correction, modification, and shutdown; no scale-robust solution has been demonstrated.
- Compute governance. Jurisdiction-specific export controls, infrastructure monitoring, and thresholds used as imperfect proxies for capability—not as a definition of AGI or safety.
Anthropic's chief executive made the case for the first of these in his 2025 essay on the subject:
Dario Amodei argues that researchers still lack a precise account of why models choose particular outputs.
The Urgency of Interpretability, Primary source
Institutional levers
The coordination problem
International coordination matters because unilateral controls can change both safety incentives and competitive position. The policy landscape is no longer entirely voluntary: the EU AI Act imposes binding duties on providers in scope, while lab safety frameworks and summit commitments remain largely voluntary. No binding global frontier-model arms-control treaty currently exists.
An early international starting point was the non-binding declaration endorsed by 28 countries and the EU at the first global AI Safety Summit:
28 nations and the European Union commits signatories to AI that is designed, developed, deployed and used safely and responsibly.
GOV.UK, Signed statement
- Risk-proportionate pre-deployment evaluation for dangerous capabilities before unrestricted release.
- Incident reporting that lets regulators and developers learn from serious failures and near misses.
- Liability rules that account for control, knowledge, causation, and the ability to prevent harm.
What an individual technologist can do
Readers can influence policy as voters, workers, researchers, customers, investors, and members of civic organizations. Practical options include contributing to or funding interpretability and alignment research; pushing safety practices inside your own organization if you deploy AI systems in production; supporting the reporting and evaluation infrastructure above rather than treating it as regulatory friction; and checking the source and evidence behind persuasive claims rather than assuming either automated manipulation or genuine consensus.
Research Policy Organizational
In this chapter
Scalable oversight: supervising more capable systems The technical approaches researchers are testing to let human overseers supervise AI systems more capable than themselves, as tested through AI debate.
Source-checked 5 min
Compute governance: chips, data centers and thresholds How chip export controls, data-center reporting and compute thresholds can govern advanced AI development, including their limits and enforcement challenges.
Source-checked 5 min
International treaties and summits: where governance stands A plain accounting of the Bletchley Declaration, follow-on AI safety summits, and how far international AI governance has gotten toward a ratified treaty.
Source-checked 5 min
Corrigibility: systems that accept correction Why "just build in an off switch" is harder than it sounds, and what corrigibility research is trying to solve—including why an agent may resist shutdown.
Source-checked 5 min
Whistleblower protections inside AI labs How whistleblower rights, reporting channels and anti-retaliation rules can help AI-lab employees expose safety failures before they become public harms.
Source-checked 5 min
The open-source AI debate: safety tool or safety risk Examine the strongest safety, research, competition and misuse arguments for and against releasing powerful AI model weights openly, including a $200 jailbreak.
Source-checked 5 min
AI law and enforcement in 2026 What AI rules are binding as of September 2026 across the EU, US, and international frameworks—and what remains voluntary and where enforcement bites.
Source-checked 6 min
Frontier evaluations, safety cases, and incident reporting How capability tests, safeguard evaluations, structured safety arguments, and incident systems can turn frontier-AI promises into auditable evidence.
Source-checked 6 min
Frontier AI labs' safety frameworks, compared Six labs' safety frameworks compared: thresholds, safeguards, and what happens when one is crossed, checked against independent audits of actual practice.
Source-checked 12 min
AI regulation, country by country: where the rules actually stand A September 2026 country-by-country snapshot: the EU's delayed AI Act, the US state patchwork, the UK's no-statute stance, China's labeling rules, and more.
Source-checked 7 min