Do AI delegates exploit their manual-playing counterparts?

Beyond the headline welfare comparison across modalities, the paper reports a specific distributional finding inside the Delegate condition: "unassisted counterparts (non-adopters) in Delegate games actually achieved a higher average individual surplus (0.218) than the active AI adopters who utilized the Delegate (0.201)." The paper reads this as "an instance of transactional equity: rather than acting as a predatory tool that exploits counterparts, the Delegate agent upgrades the shared decision space, permitting counterparts to capture equal or even greater welfare gains." It frames the result as "an empirical example against the central ethical concern of algorithmic exploitation in multi-party systems: that an optimization agent might maximize joint surplus by extracting value from unassisted counterparts through highly asymmetric splits."

The mechanism offered is market-wide spillover rather than direct transfer: the Delegate acts as a local "market maker, introducing higher-quality, non-conservative proposals into the shared offer pool, lifting the welfare floor for adopters and non-adopters alike." The paper's trade-level decomposition supports this: "manual proposals in Delegate-access games yield higher joint surplus (3.110) than those in Advisor (2.678) or Coach (2.746) games (p< 0.01)" — even the trades the Delegate never touches become more valuable once a Delegate is present in the game, because counterparties are responding to a pool of offers the Delegate has reshaped.

This bears on Can AI agents designed by competing principals destroy bargaining surplus? from the opposite direction: that account's open conjecture is that letting principals tune agents' incentives against each other can destroy surplus through escalating toughness, while this finding is a concrete case where a single, fixed, cooperative Delegate raises the floor for everyone in the game, including players who never adopted it. Read together, the two notes suggest the exploitation risk in multi-party AI delegation depends on what the delegated agents optimize for and against whom, not on delegation as such. It also extends Do people prefer AI advisors even when delegation works better?, the sibling note on this same experiment, by showing the Delegate's welfare benefit is not confined to its own adopters.

The finding rests on one chip-bargaining game, one LLM, and a comparison of group averages rather than matched counterfactuals for individual non-adopters, so it does not establish that delegation is generally non-exploitative — only that, in this setting, a capability-matched Delegate happened to lift rather than depress the outcomes of players who stayed manual. Whether the same spillover would hold if the Delegate were optimized purely for its own adopter's surplus, rather than behaving as a generally cooperative proposer, is left untested.