Frontier AI Outscores All 12 CPAs in Mercor Accounting Study, 18 Months After Lagging
MFordFuture · x · 2026-10-03
Mercor hired 12 licensed CPAs (avg 5.5 years experience) to complete simplified APEX-Accounting tasks. Frontier models were faster and more accurate than every accountant, including the best one. Just 18 months ago, the best models scored below the accountants' 37% average. Authors caution the tasks skew toward AI strengths (detail-oriented instruction following) and exclude client communication, but argue accounting will change substantially even if model progress froze. The study also probes the trillion-dollar paradox of soaring benchmark scores with little measured productivity growth.
More from AGI Musings
- Kent C. Dodds says AI agents tanked his course sales and changed education forever — craigsdennis · 2026-10-24
- Trusting AI for 24/7 incident response means handing root to whoever owns the GPUs, researcher warns — jd_pressman · 2026-10-24
- Even unhackable stacks won't save you: AI attack pace vs slow software migration — jd_pressman · 2026-10-24
- AI can both adversarially attack and harden every complex system, from cyber to bio — beffjezos · 2026-10-24
- Individuals will always outpace institutions on adopting AI-rewritten secure stacks — jd_pressman · 2026-10-24
- AI safety debate: those right on AI cyberattacks may be wrong on bioweapons — basedjensen · 2026-10-24