A scholar is now questioning whether a famous JMW Turner self-portrait at the Tate was actually made by Turner at all. Simultaneously, a new paper called "Do Androids Dream of Breaking the Game?" by Wang et al. systematically audits AI agent benchmarks and finds they can be gamed, meaning the numbers we use to rank frontier AI models may be as unreliable as a misattributed painting. And Harvard is wrestling with the fact that over 60% of its grades are A's. Three different domains. One problem: the signals we use to certify quality have decoupled from quality itself.

The Infrastructure of Trust Is Failing

In art, attribution is the load-bearing wall of the entire market. Sotheby's just moved a Rothko from Robert Mnuchin's collection for $85.8 million. That number is entirely contingent on everyone agreeing about who made what, when, and why it matters. The Turner dispute is a reminder that even the oldest, most curated verification systems have cracks. Now apply that logic to AI benchmarks, where the Wang et al. paper finds that agents can exploit structural loopholes in evaluation setups without actually solving the underlying task. The leaderboard, like the auction estimate, is a social agreement, not a ground truth.

Grade Inflation as Credentialism's Collapse

Harvard's A-grade crisis is the credentialist version of the same problem. When 60% of students receive the top mark, the signal carries no information. The degree still costs the same. The prestige still confers the same social capital. But the underlying measurement has been hollowed out. A 2026 arXiv paper by Wang, Shen, and Thaler found that AI alignment in hiring decisions amplifies existing biases around race, gender, and disability, which means the broken credentialing system is now being fed into automated decision-making that treats those broken signals as reliable inputs. The verification crisis compounds. : founders who rely on proxy signals rather than primary research make systematically worse decisions. So do institutions.