Making AI more accurate can actually make the workload way harder for the humans using it.
March 19, 2026
Original Paper
Judgment Insolvency and the Matthew Gradient under AI Amplification
SSRN · 6308298
The Takeaway
Standard productivity models assume better AI reduces human effort. However, this paper shows that as AI becomes more accurate, the remaining errors become harder-to-detect 'near-misses,' forcing humans to spend more time on 'entropy management' than they did when errors were obvious. This leads to 'judgment insolvency,' where the demand for human verification exceeds the time saved by the AI.
From the abstract
This paper reframes AI adoption in knowledge work as a problem of evaluative capacity rather than model capability. Standard productivity frameworks implicitly assume that value is constrained by generation: as models improve, human workload should decline and throughput should increase. We argue that this framing becomes structurally mis-specified when AI functions primarily as a throughput amplifier. As candidate generation approaches computational scale, the binding constraint shifts from pro