Proposed 'comprehension audits' would pause AI work and dependent releases if humans can't explain it

A paper coauthor proposes 'comprehension audits' to test whether humans understand key AI R&D contributions. Independent auditors would select work for the responsible people to explain; a failed audit would pause the contribution and any dependent release. The coauthor says their analysis of open-source AI repositories found that human review comments per line of code roughly halved from early 2024 to mid-2026. The authors seek a pilot with a frontier lab.