Generative AI has broken the assumption that online assessment performance reflects student knowledge, and the traditional countermeasures-proctoring software, AI detectors, lockdown browsers-are expensive, adversarial, and increasingly ineffective. This paper proposes a different paradigm: random post-assessment audits with credible penalties, modeled on the sampling-based enforcement strategies used in tax auditing, financial regulation, anti-doping, and nuclear nonproliferation. A game-theoretic analysis shows that with a penalty of 200 points (on a 100 point assessment) and a two-audit appeal mechanism, an instructor needs to conduct only 2-4 brief conversations per assessment in a class of 30 to make cheating irrational. The false positive rate-not the audit rate-is the critical parameter that determines fairness to honest students, and a two-audit appeal reduces the effective false positive rate to levels where a modest point bonus makes honest high-performers net beneficiaries of the system. An accompanying interactive simulator allows instructors to calibrate parameters for their own courses. The paper concludes with a call for small-scale pilot studies to empirically validate the framework before broad adoption.