The Race — Building Democracy in the Age of AI. People race toward a shared finish line.
The Race
← Trends

Supporting case

An economic model explains why more detected AI violations need not mean worse compliance

What happened

Joshua Gans and Richard Holden's paper, When Does Randomized Oversight Align AI Agents That Can Conceal?, models an agent's incentives to break rules and hide evidence. The full September 29 version is accessible; NBER lists the paper for October, without an established exact publication day. This is theoretical research, not an experiment measuring current agents or a quantified finding about deployed systems.

The model distinguishes preventing misconduct from detecting it. Stronger checks can deter some violations while inducing the remaining violators to conceal more effectively. Unpredictable audits can work when evidence survives, penalties are sufficient and the agent cannot anticipate the audit draw. Erasable evidence weakens that arrangement. Incentives to stop an unsafe task and loss of future access can also matter. The paper does not recommend generally weakening audits: lower detection can mean either better compliance or better concealment. Full paper ↗.

For The Race, my inference is that oversight institutions need evidence-preservation and incentive design as well as an audit rate. Counting detected breaches alone can misdescribe accountability. The findings do not demonstrate a new regulator's authority or a successful real-world oversight institution.

Dates and implementation stage

September 29 arXiv full version; NBER October listing, exact day unknown

Supporting sources

Related trends