Looking for Bidding Teammates: Why AI-Assisted Reviewing Makes Stranger Collusion Sustainable

Reviewers are recruiting collusion partners in public. On open social platforms, strangers post offers to trade bids on each other's submissions and return inflated scores, an arrangement that requires no prior relationship and no existing ring. We ask why it is stable. Modeling recruitment as a four-stage game between strangers, we show that the reviewing stage is a Prisoner's Dilemma for every parameter setting, so the one-shot game has a unique subgame-perfect equilibrium without collusion and the arrangement should unravel. It survives only through repetition, and the binding constraint is then the effort of writing an inflated review that survives discussion. That effort is what AI assistance removes. The sustainability threshold is increasing in effort cost, so lowering it enlarges the set of self-enforcing partnerships with no change to detection, sanctions, or the value of publication. In a calibrated conference simulation the share of reviewers who can sustain collusion rises from $68.2\%$ at hand-written effort to $95.2\%$ when the submitting authors draft the review, and the optimal ring grows from $4.6$ to $8.6$ partners. Meanwhile no detector we test exceeds $F_1 = 0.322$ against an attacker who camouflages bids, distributes them around a ring, and manipulates affinity. AI-assisted reviewing is therefore not only a question of text quality: it changes which collusive arrangements are self-enforcing.