Recently, I've been feeling very hopeful about the idea that surging interest in AI auditing and evaluation could be an entry point to solving AI's serious issue with unclear "data rules". At serious risk of having fallen into the "thing I've been thinking about a lot is applicable to everything" trap, I think auditing and evaluation -- and recent efforts to support more evaluator institutions -- can actually be tied to a pretty effective data dividends scheme as well.

The idea would be to implement a "data dividend" that, rather than being funded by a tax on compute or automation, could be a tax on "unaccounted for capability".

It would be a "presumptive commons-rent tax on frontier AI".

We could presume that any monetized, highly capable system derives some share of its value from commons data and tax that share by default. Firms can "rebut the presumption" by showing receipts: evidence of provenance, acquisition, and ablations showing that the relevant capabilities actually came from accounted-for data. More capability -> higher tax rate without rebuttal; more explained and accounted for data pushes the liability down toward zero. The safety/audit ecosystem (that we need to build anyway!) could do the verification.

Ideally this could create an avoidable Pigouvian rent tax -- so either it has lots of funds (napkin math important here as always) OR it creates incentives for the entire frontier ecosystem to run on highly documented data from ecosystems where data creators have leverage.

Of course, could be designed to complement other short term factor taxes and related schemes.

Longer post to come on this!