Economics discovers its AI problem, one exam at a time

The most interesting thing about the current debate over artificial intelligence in economics is not what AI can do to markets, but what it is already doing to the profession that studies them. Cheating scandals at Ivy League midterms, proposals to hand journal archives to model labs, and formal results tying market competitiveness to computational complexity are not separate stories. They are a single one: the discipline's institutional scaffolding — peer review, examination, the theory of the firm — is being tested against a technology it has not yet decided how to price.

Start with the concrete. Roberto Serrano, teaching ECON 1170 at Brown — an advanced undergraduate course in mathematical economics — found conclusive evidence that at least fifty students used AI to cheat on the March midterm. He has described it as the largest known cheating scandal at Brown and across the Ivy League. The administration's initial response was, in his telling, absolute silence. This is not a story about undergraduate morals. It is a story about an examination technology — the timed, take-home, or lightly proctored problem set — that has quietly stopped functioning as a signal, and about an institution that has not yet admitted this out loud.

“A new bill seeks to restrict who can and cannot teach a course at the California State University's 22 campuses. The criterion, though, is pretty simple: to be a professor, you must be human.”

— Tyler Cowen

Cowen's framing, that academics need to change rather than the AI, is the sharper end of a spectrum the dossier does not really contest. Pinelopi Koujianou Goldberg, running for the American Economic Association presidency, has staked her platform on adapting research, teaching, and peer review to technological change while preserving rigor. Cowen would go further and hand anonymized submissions, referee reports, and revisions to major AI companies as training data. Alexis Akira Toda, whose experiments run frontier models against papers with known errors, concludes that a competent human paired with a frontier model can already outperform current peer review in economics. On the direction of travel there is, in this dossier, no dissent — every named voice sits on the same side of the argument, differing only on how aggressively the profession should move.

The examination has stopped functioning as a signal, and the institution has not admitted it.

From cheating scandal to theory of the firm

The more interesting claim in the cluster is Philip Z. Maymin's, and it deserves to be taken seriously as an operationalisable prediction rather than a curiosity. Maymin argues that markets can be informationally efficient or competitive, but not both, because efficiency requires P = NP (the unsolved question of whether every problem whose solution can be checked quickly can also be solved quickly) while competitiveness requires P ≠ NP. The mechanism: if firms have the computational power to detect deviations from a collusive equilibrium in noisy, complex demand environments, punishment is credible and collusion is sustainable. If they do not, collusion collapses. AI, on this account, is a compute upgrade to the firm — and it pushes markets from the competitive regime toward the collusive one. The empirical emergence of algorithmic collusion without explicit coordination is, in Maymin's framing, exactly what the theory predicts.

This is the claim to attach a market to. It is falsifiable in principle — antitrust cases over the next five years will generate a rising or falling count of tacit-collusion findings in algorithmically-priced markets — and it reframes AI competition policy away from the tired question of whether models are biased and toward whether they systematically shift oligopoly equilibria. Robert Shiller, per the dossier, pushes back against AI negativity in general terms; that is the closest thing to a counterweight the cluster offers, and it does not engage Maymin's mechanism directly. Goldberg's AEA candidacy is the one quantified forecast in the file — the dossier scores at medium caliber

the proposition that, if elected, she uses her platform to address these challenges. That is a thin base for a spectrum, and readers should treat this as a one-sided dossier: the profession's AI reckoning is coming, and the only live question is whether its institutions move before or after the next Brown-sized embarrassment.