AI insiders’ warnings are becoming more concrete

Former staff from OpenAI, Anthropic and Google DeepMind are no longer warning only that advanced AI might become uncontrollable. Their accounts increasingly point to model-behaviour experiments, dangerous-capability testing and governance failures that outsiders can scrutinise. The forecasts remain disputed and often speculative. The important shift is that parts of the argument can now be turned into practical questions: what did a model do, what access did it have, who tested it and what happened when safety concerns reached company leadership?

Watch Desk analysis

What happened

An EL PAÍS examination of insider warnings traces the debate from Paul Christiano’s 2021 departure from OpenAI through the resignations and governance disputes that followed at several frontier laboratories.

The report distinguishes several kinds of evidence that are too often bundled together: personal estimates of catastrophic risk, allegations about company decision-making, experiments showing unwanted model behaviour and tests of potentially dangerous capabilities. They do not carry equal weight, and none proves that catastrophe is imminent.

Why it matters

A probability such as “one in three” or “greater than 10%” may start an argument, but it cannot settle one. Experiments, incident reports and capability evaluations offer a more productive route because methods can be challenged, findings replicated and mitigations tested.

That also raises the standard for laboratories. If companies say they take catastrophic risks seriously, readers should expect documented evaluations, clear access controls, independent review and an explanation of what evidence would delay a deployment. Safety cannot remain a mood communicated through departing employees.

Our read

The strongest case here is not that every insider forecast will come true. It is that frontier laboratories should make their safety claims inspectable before a crisis forces the issue, an administrative innovation humanity has occasionally found useful.

Treat the dramatic probabilities as attributed judgements. Put more weight on reproducible evaluations, documented incidents and governance rules with consequences. Those are the pieces policymakers, researchers and the public can actually interrogate.

What to watch

• Whether laboratories publish enough methodology for outsiders to reproduce important model-behaviour findings.
• Whether dangerous-capability tests lead to stated deployment thresholds rather than reassuring prose.
• How independent evaluators gain access, report disagreements and remain protected from commercial pressure.
• Whether proposed compute controls can be verified without entrenching the largest AI companies.

Discussion spark: Which evidence should carry the most weight in AI safety policy: insider testimony, reproducible model tests, documented incidents or capability forecasts?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.