AI and the Erosion of Self-Selection as a Social Norm in Science

How technology affects a community depends on the social norms it inherits, and these in turn affect what technology can do. In science, a critical social norm is author self-selection: the shared restraint by which authors withhold work that does not merit review, keeping submission volume tied to genuine effort. AI is, above all, a generative technology: it generates candidate work faster than peer review can assess it. In the context of this asymmetry, self-selection is key to a healthy peer-review ecosystem. We use game theory to understand whether authors keep self-selecting as generation cheapens while review stays limited and imperfect, and whether directing AI at review can hold the norm in place. We find that self-selection holds only while the intensity of AI in manuscript production stays below a threshold, past which the submission of weak work becomes profitable and the norm gives way to indiscriminate submission. This does not reverse: we prove indiscriminate submission, once reached, is self-sustaining, and self-selection cannot be restored by reducing the intensity of AI in manuscript production. Targeting AI at review to close the generation--verification gap cannot reverse it either: adding reviewing capacity does not supply the accuracy that decides whether weak work is published. Even a maximally accurate reviewer does not restore the norm in our calibrated model; only added throughput combined with near-complete correction of AI's accuracy loss does. Technology and the norms it inherits shape each other, and so must evolve together: AI that outruns the institutions sustaining scientific trust erodes them beyond what more AI can repair.