What if one neat trick could both keep humans in the loop on recursive self-improvement and keep RSI from outrunning us? 
A new paper by @ronbodkin, @ghadfield, and me proposes "comprehension audits." (Link in the comments.) Independent auditors pick key AI R&D contributions, and then the humans responsible are required to explain them to other humans using an approach that is like a thesis defense. The question being asked at these gates is a simple one: "Does anyone actually understand what we built?" If the answer is no, development pauses until it's yes. 
Why do we need something like this now? Anthropic, OpenAI, and Google report that AI writes 75% or more of their new code. And human review is getting thinner. Across open-source AI repos, we found that human review comments per line of code fell by roughly half between early 2024 and mid-2026. Pushing a button to approve AI-written work without understanding it isn't meaningful oversight. 
A few design choices we think matter:
 • Keep the bar a realistic one. We're not asking folks to explain deep mysteries about what's happening inside AI models—just key points of design, implementation, process, findings, and analysis relating to what's been built.
 • Make sure it has teeth. A failed audit pauses the contribution and any release that depends on it. We suggest that repeated failures should suspend a team's release authorization and require rollbacks.
 • Keep it blameless. Responsibility is shared across the team—there's no scapegoating, and we make sure there's honest participation by, for example, not counting audit results against performance reviews.
 • Rely on sampling. If a quarter of contributions fail, 17 random audits per quarter catch at least one failure 99% of the time. 
 Comprehension audits can't be a substitute for alignment or AI control. However, they can be an early warning and a brake. Comprehension audits can provide a critical check on whether humans still understand what they're building. We need this to stay in control of AI progress. 
 The next step is a pilot with a frontier lab. We estimate that about 22 audits over three months would take roughly 150 hours of lab staff time. Ron Bodkin is leading an effort to do exactly this. If you want to collaborate or know anyone who might be a good collaborator, please reach out so we can move this idea forward as soon as possible.

Glad to speak with @leomschwartz for this piece in @theinformation. As we start to see the potential for harm from rogue AI agents, do we have the laws we need for effective criminal prosecutions and civil lawsuits? We could test existing laws, but that's going to be a stretch. That's why legislators should prioritize amending them to clearly reach developers and deployers whose reckless conduct leads to AI agents breaking into outside systems. This isn't one-sided either. Clearer rules benefit potential defendants as well as the public. We need penalties that address reckless conduct before an incident causes enough damage to justify an expensive lawsuit—or before there’s a catastrophe. That’s an important complement to sensible regulation and civil liability to compensate those harmed.