Future Matters finds 13 of 22 AI donors can't measure governance grant impact

A LessWrong linkpost argues that AI governance funding suffers from dangerously weak monitoring. Most donors never check whether their grants worked: they model impact before sending money but rarely verify it. The fix is simple: agree written indicators, ask grantees to assign probabilities, and score them yearly. Donors usually get reports, but unless they ask for specifics, the reports omit the most useful information. One client’s report touted 20 policy recommendations, but when pressed, the grantee admitted only 20% had been even partially implemented. Berlin think tank Future Matters found that 13 of 22 major earning-to-give donors couldn’t tell whether AI governance work achieves anything, including four out of six who work at frontier AI labs. The piece calls this "illegibility" a choice donors make by not asking the right questions at the right time. Monitoring is often most rigorous where uncertainty is lowest, like bednets, and absent where it's highest, like policy and advocacy. Even GiveWell, the field-leader in rigour and transparency, has conducted very few retrospective evaluations, only starting that process in 2025.