Agentic Interactions is a December 5, 2025 working paper by Alex Imas (UChicago Booth), Kevin Lee (Michigan Ross), and Sanjog Misra (UChicago Booth). It is an experimental study of whether human heterogeneity persists when economic bargaining is delegated to AI agents, and reports that AI-mediated outcomes are more dispersed than human-to-human outcomes rather than converging. The paper introduces two framework terms, "machine fluency" and "specification hazard."
Research question
The paper asks whether human heterogeneity persists, or even amplifies, when economic decisions are delegated to AI agents. The authors note that standard economic intuition, and much frontier-lab marketing, predicts homogenization: identical models maximizing identical objective functions should converge. The paper argues the opposite and presents what the authors describe as the first large-N experimental evidence on the question.
Experimental design
The study uses a multi-round used-car bargaining game under the induced-values methodology of Smith (1976, 1982), which fixes valuations by construction so that any observed heterogeneity is non-instrumental rather than a matter of differing tastes. The Buyer's outside option is a $22K dealer price and the Seller's outside option is an $18K trade-in, leaving a bargaining surplus of $4,000. Participants write Buyer-side and Seller-side prompts for AI agents, which then negotiate against other participants' agents. Each Buyer prompt was paired with each Seller prompt in a round-robin tournament, and outcomes were averaged. The primary agent model was OpenAI's gpt-4.1-mini, with robustness checks across the larger GPT-4.1 and the smaller GPT-4.1-nano.
Participants were recruited via Prolific and restricted to the US, UK, and Ireland. The sample comprised N=299 in the AI-mediated condition and N=304 in the human-to-human condition. Heterogeneity was decomposed using fixed-effects regression and a Shapley-value decomposition of R². The authors report that robustness across model sizes indicates the dispersion is not an artifact of stochasticity in any one model.
Findings
Heterogeneity persists and intensifies
The paper reports that 73% of variance in AI-mediated bargaining outcomes is explained by individual fixed effects. Because instrumental preferences are pinned down by the induced-values protocol, the authors argue this dispersion can only flow through the prompt, that is, through the non-instrumental characteristics of the human principal that get encoded in it. AI-mediated outcomes have 16.5% higher variance than human-to-human outcomes under identical bargaining parameters and interface. The 50/50 fairness norm appears in 34.7% of human-to-human deals but only 14.3% of AI-mediated deals; the authors interpret this as implicit social norms attenuating when bargaining is delegated to agents, removing a coordination device that disciplines extreme outcomes.
Demographic effects propagate and sometimes reverse
The agent had no access to the principal's demographics, so the only channel is the prompt. Despite this, the paper reports that the gender gap in negotiation reverses for Sellers under AI mediation: in human-to-human bargaining women extract less surplus as Sellers, consistent with the existing literature, but under AI mediation female-prompted Seller agents do better than male-prompted ones. Returns to education increase for Buyers under AI mediation, with no such effect human-to-human. Personality (Big Five), social preferences (Dictator, Prisoner's Dilemma, and public goods games), and game-theoretic behaviors explain substantial shares of outcome variance via the Shapley decomposition.
The Shapley decomposition of explained variance by predictor block:
| Block | Shapley R² | Share of total R² |
|---|---|---|
| Demographics | 0.047 | 27.6% |
| Game-theoretic behaviors | 0.037 | 21.6% |
| Negotiation experience | 0.024 | 14.0% |
| Big Five personality | 0.023 | 13.5% |
| Specific bargaining traits | 0.015 | 8.7% |
| Effort (time/iterations) | 0.014 | 8.5% |
| Cognitive Reflection (CRT) | 0.006 | 3.3% |
| Risk and patience | 0.005 | 2.9% |
Buyer/Seller asymmetry
The paper reports that humans are systematically better at instructing AI agents as Buyers than as Sellers. The authors conjecture this reflects asymmetric experience, since most people buy used cars more often than they sell them. Sellers' agents leave more variance unexplained, which the authors suggest may indicate that humans have less articulable knowledge to encode in the Seller prompt.
Framework concepts introduced
The paper introduces two terms. Machine fluency is the ability to instruct an AI agent to align effectively with one's objective function (machine fluency). The paper reports that machine fluency varies systematically with principal type, correlating with cognitive sophistication, education, prior experience, and personality, and that it predicts economic outcomes. The authors propose machine fluency as a new source of economic inequality in agentic economies, which they analogize to literacy and numeracy gaps in the LLM era.
Specification hazard is presented as an informational asymmetry distinctive to principal-(AI)agent relationships (specification hazard). In classical principal-agent theory the principal worries about the agent's hidden information or hidden action; with AI agents, the authors argue, the principal may instead be unable to specify what they want with enough precision, while the prompt inadvertently injects the principal's priors, biases, and personality into the agent's objective function. Misra (2025) provides the theoretical companion.
The authors position machine fluency as a candidate axis of the AI divide, alongside access, infrastructure, and skill, arguing that the ability to write effective agent prompts will shape distributional outcomes. They frame specification hazard as an update to the principal-agent problem applied to AI, with outcomes loading on principal characteristics the agent cannot see because the prompt is a noisy carrier of the principal's preferences.
Provenance
Working paper, dated December 5, 2025, authored by Alex Imas and Sanjog Misra (UChicago Booth) and Kevin Lee (Michigan Ross). The paper cites Shahidi et al. 2025 for the demand-side framing; see The Coasean Singularity? Demand, Supply, and Market Design with AI Agents — Shahidi, Rusak, Manning, Fradkin, Horton (2025), a companion theoretical chapter that is co-cited and cross-referenced.
Relationships
- supports: Principal-Agent Problem Applied to AI — provides an empirical mechanism (prompt as noisy preference contract)
- supports: Agentic AI — quantifies how heterogeneity persists in agent-mediated markets
- supports: Machine Fluency (new) — original empirical anchor
- supports: Specification Hazard (new) — original empirical anchor
- supports: AI Divides (Literacy / Occupational / Ethico-Philosophical) — adds machine fluency to the list of AI-driven inequalities
- related: The Coasean Singularity? Demand, Supply, and Market Design with AI Agents — Shahidi, Rusak, Manning, Fradkin, Horton (2025) — companion theoretical chapter co-cited and cross-referenced (the Imas et al. paper cites Shahidi et al. 2025 for the demand-side framing)
- contradicts: the homogenization prediction implicit in much frontier-lab marketing (and in Bertrand-competition arguments about internet markets)
- related: Alex Imas (new), Sanjog Misra (new)