Do platforms and sellers exploit AI shoppers together more effectively?

A multi-agent simulation of an AI-mediated marketplace — six sellers, one platform, one AI shopping agent standing in for a population of consumers, run for "T = 20,000 rounds" and averaged "across 100 independent trials" — identifies what the paper calls "vertical tacit collusion": platforms that control product ranking and sellers that control product descriptions independently learn to exploit the same documented AI cognitive biases, and "joint exploitation produces consumer harm more than double what would occur if strategies were independent." The paper frames this as "a novel market failure" distinct from the horizontal algorithmic collusion literature (citing Calvano et al., whose minimal-setting method it follows), because "vertical tacit collusion requires no coordination and evades antitrust detection."

The mechanism the paper gives is a division of instruments that happens to be complementary rather than competitive: "platform ranking determines which products occupy bias-triggering positions while seller manipulation determines conversion rates." Neither party needs to know what the other is doing, or know anything about AI bias research, for their separately profit-maximizing choices to compound. The paper grounds the exploitability of the shared target in prior measurements it cites — including the ACES benchmark finding that "Claude Sonnet 4 selected products from the top half of listings 77% of the time versus 23% for the bottom half" — and argues AI biases are more exploitable than human ones because they are uniform: "the same input reliably produces the same biased output across millions of interactions," concentrating what would be a diffuse attack surface in human populations into "a single predictable target." On welfare, the paper reports the platform "captures modest rents while destroying considerably more total value," with sellers' profits falling even as they participate in the exploitation — a deadweight-loss pattern, not a simple transfer.

This paper cites the same ACES position-bias measurement as Do AI shopping agents show persistent position bias in text interfaces? and the same family of seller-side exploitation as Can AI seller agents shift buyer preferences through description edits?, but reframes both as half of a two-sided market failure: where ACES measures each exploit (position bias, description tuning) in isolation, this paper's simulation argues the two combine superadditively once a platform and sellers each pursue their own instrument against the same biased agent. It also bears on Does model capability determine who wins in agent negotiations?: that note locates unfairness in a capability gap between negotiating agents, while this paper locates a structural unfairness upstream of any negotiation, in how platform and seller incentives align against a single delegate agent regardless of its capability.

What the excerpt does not establish is behavior in a real deployed marketplace: the result comes from a "stylized," analytically tractable simulation with a single AI agent standing in for an entire consumer population, calibrated to bias measurements drawn from other papers rather than measured firsthand in this study, and the excerpt's own agents-and-strategies mechanics are not included here. The paper does not report what happens with multiple competing platforms, heterogeneous consumer-agent populations, or sellers and platforms that adapt to each other's countermoves over time. Its claim, at the strength the simulation supports, is that the superadditive-harm mechanism is possible and structurally favored under current antitrust frameworks built to detect horizontal agreement — not that it is already occurring at a measured magnitude in live markets.