Anka Reuel | @ankareuel.bsky.social

Now’s prob a good time to mention that I’ll start as an Assistant Professor with appointments at @Kennedy_School and @HarvardEngineer in 2027, working on… AI evals! My lab will also work on monitorability, incident analysis, and verification to advance technical AI governance 1/

Way to go @HarrisPolicy!! I’d love to see more universities open faculty positions at the intersection of AI & policy (& follow great schools like @Kennedy_School and @JohnsHopkinsSGP who had such hiring rounds last year!)

Love this for New York State 🩷

+1. One of the most underestimated problems imo with calls for more transparency and independent evaluations is that disclosure alone doesn't achieve much. for independent evaluations to function as effective oversight mechanisms, we need to think beyond who runs them and who gets access to the reports. There are two issues here we need to address:

First, as Nat points out, we often don't know what to do with evaluation findings in the first place. Most evals we're running today lack established predictive validity, clear thresholds for when action should be taken, or even agreement on what that action should be. We need more action-oriented evaluations that work backwards from "under what conditions should we take what action?" to "how do we measure these conditions or their precursors?" rather than measuring things like RSI for the sake of measuring it and hoping someone downstream figures out what to do with the results.

Second, we need people on the receiving end who have the expertise and capacity to act on the findings. This is particularly relevant for non-public evaluations whose reports are shared only with regulators (e.g., under the EU GPAI Code of Practice, which requires signatories to share independent evaluator reports with the EU AI Office), many of whom are already understaffed.

There's obviously a tradeoff here. Restricting disclosure to authorities protects commercially and safety-sensitive information, but limits scrutiny to a small pool of experts, if they have any at all. Public disclosure allows for broader scrutiny, but creates disclosure risks and doesn't necessarily translate into enforcement.

I'm very much in favor of more independent evaluations but need to get better at both designing evaluations around the decisions they're meant to inform and ensuring that those receiving the results can actually act on them. Otherwise, they risk becoming little more than a PR exercise for frontier labs.

I often get asked what I read to keep up with everything happening in AI governance. Substack is one of my go-to places for more in-depth commentary. They just launched lists today so I made one with all my fav AI governance Substacks:

ankareuel.substack.com/lists/ai-gover…

Please suggest additions in the comments if you have one that you love and that’s not on the list yet!