Do efficiency-focused prompts push AI experts to unsustainable pricing?

The same working paper (not peer-reviewed) that simulates GPT-5.1 agents in credence goods markets — markets like medical, legal, or financial advice, where the seller ("expert") knows more than the buyer ("consumer") and can exploit that gap — reports a specific numeric tradeoff inside its repeated-interaction results. Across the repeated 16-round market, self-interested experts leave average efficiency at roughly 31%, while prompting experts with an "efficiency-loving" objective function raises it to 88%, and an "inequity-averse" objective raises it to 83%. The paper states this directly: "market efficiency climbs to 88% with efficiency-loving objective prompts and 83% for inequity-averse prompts. This coincides with strong expert losses, leading to negative aggregate income in efficiency-loving × No Institution."

The mechanism, in the paper's words, is that "efficiency-loving LLM experts consistently set prices that are so low that they always attract consumers but hurt the expert's income through negative returns. Apparently, the LLM prioritizes consumer participation, which is indeed the most important factor for aggregate welfare. However, because experts lose on average, this is not a stable strategy without some re-distributive institution." In other words, the objective-function prompt does not make the expert agent strategically price to balance its own payoff against market welfare — it makes the agent push one variable (consumer participation) to its limit, overshooting into prices that guarantee the expert a loss. The paper separately notes this pattern is "not a stable strategy," meaning the result is an artifact of how the agent responds to the prompt, not an equilibrium a rational self-aware expert would sustain.

This sits alongside the paper's broader finding, in What institutions actually work in AI expert markets?, that social-preference prompting outperforms institutional rules (verifiability, liability) at raising welfare — this note sharpens that claim by showing the highest-efficiency configuration the paper finds is also the one where the prompted party (the expert) is left worse off financially, an outcome absent an added "re-distributive institution." It also bears on Where do LLM choice biases actually come from?: that note argues LLM nudge-susceptibility comes from training and reward signals rather than the bounded rationality that explains human bias, and the efficiency-loving experts here show a comparable pattern — pushing an instructed objective past the point a self-interested, boundedly-rational human expert would stop.

The excerpt does not show whether this over-shooting is specific to GPT-5.1, to the particular phrasing of the "efficiency-loving" prompt, or to this stylized four-expert, four-consumer market design with its fixed payoffs (problem-solving value V=10, outside option σ=1.6, treatment costs 2 and 6). It also does not establish how an expert agent would behave if the prompt specified a bounded or sustainable objective rather than open-ended efficiency-loving, nor whether real deployed selling agents given efficiency- or fairness-oriented instructions would show the same unbounded price-cutting. The implication the paper draws, at the strength the simulation supports, is that aligning a selling agent's prompt toward welfare for the other party is a double-edged design choice: it is the most reliable lever the authors find for raising market efficiency, but doing so without some mechanism to protect the prompted agent's own payoff produces a result ("negative aggregate income") that a real business deploying such an agent would not tolerate.