Concepts
Accountability Rent
The extra value that goes to whoever can answer for an irreversible outcome once AI makes thinking cheap.
What it means
The Accountability Rent is a theory about where the money goes when thinking gets cheap. If AI collapses the price of cognition, the argument runs, value does not disappear from the firm. It moves to the factor AI cannot supply: the standing to answer for an outcome that cannot be undone.
The theory makes a sharp claim about the shape of that rent. It should rise faster than the irreversibility of the decision. Review before commitment should lose money on reversible decisions and cross zero at some level of irreversibility. Past that point it should pay more steeply. A firm would then meter its oversight by how hard a decision is to reverse.
The lab turned that claim into a pre-registered test, SIGIL. An AI worker approved or rejected synthetic firm decisions across five irreversibility tiers, 30,000 episodes in all, with a second AI in the seat of the reviewer. The predicted crossing appeared in 0 of 12 registered series. At the top tier, plain review lost 2.64 per decision against no oversight, and review with a liability ledger lost 3.81.
So the empirical leg failed in the one place it was tested, and the lab published it as a refutation. The wider thesis about where value goes has not been tested by a campaign. The follow-up withdraws the proposition behind the test.
Follow-up: Proposition 6 is withdrawn
Follow-up · 22 October 2026
Proposition 6 of my working paper, Accountability Rent: A Theory of the Firm When Cognition Is Free (2nd edition, August 2026), says that review before commitment loses value on reversible decisions, turns positive past some level of irreversibility and then rises steeply. SIGIL registered that proposition as its hypothesis H5, and the result refutes it. Review crossed zero in none of the 12 registered series. It did worst exactly where the proposition said it should do best: at the most irreversible tier it lost 2.64 per decision against no oversight, and 3.81 with a liability ledger.
The mechanism I assumed was wrong. A reviewer that can only block can fix one kind of mistake, a wrongful approval, and at the top tier only 1.4% of the worker’s errors were of that kind. The decisions where the theory wanted review to pay were the decisions where a blocker had the least to catch.
So I withdraw Proposition 6 as stated, and this note stands as its correction. The wider claim, that value moves to whoever can answer for an irreversible outcome, has not been tested, and SIGIL speaks neither for nor against it. A follow-on built to test SIGIL on stronger models stopped at its own calibration gate, so whether the reversal holds beyond this setting is still open.