Aalok Mehta
Director of the Wadhwani AI Center at CSIS, where he leads research on AI policy.
“This is a potentially enormously consequential development in AI technology. It is happening in the private sector in ways that are very, very inaccessible for anyone but people who work in that company, and probably also only a subset of people within that company. And it seems like this is the kind of thing that the government should be able to have insight into, so that agencies and lawmakers can prepare appropriate policy responses.”
The AI Policy Podcast · Geopolitics & World Affairs · November 2026
Context
The hosts had noted that METR said much of Anthropic's data was redacted, so the public must rely on METR's judgment. Mehta argues this exposes a gap in government visibility into frontier AI development.
Transcript
…this report is sort of speculative, is based on estimates from Meter using information that we don't have insight into. And so, they acknowledge in their report that this is not ideal, and they would have loved to be able to provide more information and that they couldn't because of the redactions.
Yeah, that was the first thing that sort of grabbed my attention was their post where they were just saying that, you know, the public's going to have to rely on us on our own judgment because of how much this has been redacted.
Yeah, I mean, I think this highlights a problem, right? Like as we've seen in the AI space, right, there's a move towards more evaluations, external evaluations, organizations coming in and doing oversight of companies. But something like this raises, I think, a bigger question, which is this is a potentially enormously consequential development in AI technology. It is happening in the private sector in ways that are very, very inaccessible for anyone but people who work in that company and probably also only to a subset of people within that company. And it seems like this is the kind of thing that the government should be able to have insight into, that it should know what is happening at the very frontier so that agencies and lawmakers can prepare appropriate policy responses. And the fact that we have to rely on some sort of voluntary arrangements between labs or AI developers and external organizations, and not only that, but sort of voluntary information releases. After redactions, suggests that the government just is not going to have all the information it needs to make good policy decisions.
Right, or at least we don't know what the information the government has is, right? I mean, maybe Anthropic is providing this information to the government, but we don't know. It's all behind closed doors. There was, I believe, a post from an Anthropic co-founder saying that they're going to come…
Yeah, and I totally understand, right? Like part of their concerns are that releasing certain kinds of information could raise safety concerns. And so I understand why they might not want that information released publicly. I think you can make a much stronger case that you should still have sort…
So kind of taking a step back here for a second, can you talk to us
More from The AI Policy Podcast
