Machine fluency is the ability to instruct an AI agent to align effectively with one's objective function through natural-language prompting. The term was introduced by Imas, Lee, and Misra in "Agentic Interactions" (December 2025), which provides the first large-N experimental evidence that the ability varies systematically across people and predicts economic outcomes when AI agents act on their behalf.
Empirical basis
The concept is anchored in a large-N (N=299) experimental marketplace in which participants wrote prompts for AI agents to negotiate on their behalf in a multi-round used-car bargaining game with induced values. The agent's instrumental objective is fixed by construction (maximize bargaining surplus given a $4,000 zone), and all participants worked through an identical underlying model (gpt-4.1-mini). Despite identical objective functions and the identical model, outcomes vary substantially across principals, and that variation predicts the principal's cognitive sophistication (CRT scores), demographic characteristics (education, gender), personality (Big Five, especially Conscientiousness and Openness), and effort indicators (time spent on the prompt, simulator iterations).
In the experiment, 73% of variance in seller-surplus outcomes is explained by individual fixed effects on the human principals, and the only channel for those effects is the prompt the principal wrote. Imas, Lee, and Misra characterize machine fluency as the latent ability that mediates this channel.
Two findings from the paper bear on how the channel operates. First, the human gender gap reverses for Sellers: in human-to-human bargaining women extract less surplus as Sellers, but under AI mediation female-prompted Seller agents do better than male-prompted ones, despite the agent having no access to the principal's gender. The authors read this as evidence that gender-correlated prompting style differs in ways that interact with the AI's response distribution, the prompt being the only channel. Second, there is a Buyer-Seller asymmetry: humans are systematically better at instructing Buyers than Sellers, which the authors suggest likely reflects that most participants have more experience buying used cars than selling them and can therefore articulate a Buyer's strategy with higher fluency.
Relation to AI-driven inequality
Imas, Lee, and Misra position machine fluency as a fourth dimension of AI-driven inequality alongside three already discussed in AI Divides (Literacy / Occupational / Ethico-Philosophical):
- Access inequality — who has paid frontier-model accounts, fast inference, and agent harnesses.
- Infrastructure inequality — who has reliable connectivity, electricity, and devices.
- Skill inequality — who can use AI productively at all, that is, digital literacy.
- Machine fluency inequality — who can write prompts that elicit aligned, high-performing agent behavior.
The authors argue this fourth dimension is qualitatively distinct from the prior three because the marginal returns to fluency are amplified in agentic markets, with small differences in prompt quality producing large differences in negotiated surplus.
Relation to specification hazard
Specification Hazard is the informational characterization of the same phenomenon: the principal-(AI)agent contract is encoded in a prompt that inadvertently embeds the principal's biases. Machine fluency is the capability characterization: higher-fluency principals encode their objectives more precisely and inject fewer unwanted biases.
Publication and reception
The paper circulates as an SSRN working paper, "Agentic Interactions," whose abstract reports "significant variation in 'machine fluency' — the ability to instruct an AI agent to effectively align with one's" objective function (Source: papers.ssrn.com). Economist Tyler Cowen featured the paper on Marginal Revolution in December 2025, quoting the machine-fluency finding as its central result (Source: marginalrevolution.com). Subsequent commentary has focused on the distributional reading — that because all participants used the same underlying model, dispersion in outcomes traces to human prompting differences, making prompt skill a candidate axis of economic inequality in agent-mediated markets.
Relationships
- instance-of: AI Divides (Literacy / Occupational / Ethico-Philosophical) — a new dimension of AI-driven inequality
- depends-on: Principal-Agent Problem Applied to AI — the prompt-as-contract framing
- related: Specification Hazard — the informational counterpart
- related: Agentic AI — what fluency operates on
- supports: Agentic Interactions — Imas, Lee, Misra (2025) — primary empirical anchor