- In the Meter–Anthropic reporting on AI's effect on Anthropic's research speed, Paul Christiano said most of what the public has is Meter stating conclusions, with the full report redacted.
- He said outside scientists cannot tell whether today's systems are aligned, and that test results are hard to read without knowing how systems were trained.
Paul Christiano, who heads the Alignment Research Center, argued on the Dwarkesh Podcast that outsiders cannot check the evidence on two big AI questions. One is how much AI is speeding up Anthropic's own research. The other is whether today's systems are aligned. He spoke as a private citizen, in an episode published on 16 November 2026.
What was said
Christiano described recent reporting from Anthropic and an outside evaluator, spelled Meter in the transcript. He said the data looked consistent with the pace of AI progress having roughly doubled over the past six to 12 months. But most of what the public has, he said, is Meter stating conclusions. It promised a full report, but "the entire report has been redacted".
He explained why this is hard to weigh. Benchmark scores, the standard tests of AI ability, are a terrible way to judge the overall rate of progress, he said. Smoother measures exist inside the labs, such as how well a model predicts text in training. Meter's claim, he said, is that those measures and the research and compute inputs tell the same story. Outsiders cannot inspect them.
On safety, he said "as independent scientists really have no idea whether the systems being produced are aligned or how you would tell." Alignment means AI acting as its makers intend. The biggest missing item, he said, is how systems are trained, because evaluation results are extraordinarily hard to interpret without it.
Why it matters
Our reading, and our extension of his point rather than something he argued: Christiano spoke about frontier AI labs, not about buying software. Still, anyone who relies on a model provider's capability or safety claims faces the same gap. He called for more public visibility into the underlying data, rather than reliance on company statements.
A buyer weighing such claims could ask three things. What will the provider disclose about how the model was trained and tested? Can anyone outside see the underlying data? Has an independent party seen the full report, not a summary?
The other side
Christiano did not say the claims are false. He called the data noisy but said it is what he would expect to see six to 12 months before superintelligence.
He also named a separate obstacle to slowing down. A lone company that pauses algorithmic progress to improve security may see a rival's secrets leak and lose market share, he said. That concerns the difficulty of slowing down, not disclosure. He added that people hold different views and the debate has not run its course. The episode did not say why the report was redacted. He sits on OpenAI's board but said he spoke only for himself.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- Dwarkesh Podcast: Paul Christiano – We could be 12 months away from AIs that can overthrow the government (2026-11-16)





