Human traders lose money in digital marketplaces without ever realizing their AI opponents are smarter than them.
April 29, 2026
Original Paper
When Agents Cannot Be Benchmarked: AI Capability Opacity, Corporate Governance, and a Research Agenda
SSRN · 6655439
The Takeaway
High-frequency marketplaces now contain a structural blind spot where human participants cannot detect superior AI capabilities. This capability opacity means that a counterparty can systematically outmaneuver a person while the person assumes they are just having a bad day. Traditional benchmarks fail to capture this gap because the AI operates within the expected bounds of the system while siphoning value. Corporate governance faces a crisis because boards cannot audit the actual advantage an agent holds over a competitor. Market fairness is becoming a ghost as hidden algorithmic edges replace visible skill.
From the abstract
Sixty-nine Anthropic employees delegated a week of buying and selling decisions to AI agents in a controlled marketplace experiment (Anthropic, 2026). The agents differed only in capability. Principals matched with the more capable agent earned $2.68 more per deal as sellers and paid $2.45 less as buyers; the principals on the losing side did not detect the loss. Classical agency theory cannot account for this pattern: every action was logged, and incentives were perfectly aligned by constructio