Right-Sizing Communication and Recommendation Set Size in AI-Assisted Search

THE DISSECTION

This is a paper about optimizing the human-AI handoff interface — specifically, how much information a user should send to an AI recommendation system, and how many recommendations the AI should return. The authors treat it as a pure mechanism design problem: two costs (communication cost for the user, search cost for the AI), a Bayesian inference problem, and an optimal policy under large-dimensional preference spaces.

What the paper is actually doing: Solving for the efficient parameters of the labor-replacing process. It is, inadvertently, an engineering manual for the machine that severs the consumption circuit.

THE CORE FALLACY

The paper assumes a friction, not a termination.

It models the user as a rational Bayesian agent with costly communication and costly search, then optimizes the interaction. The entire framework implicitly assumes:

1. The human is the initiator — they "convey preference information" to the AI.
2. The AI is a servant — it "interprets the user's message" and "maximizes the user's expected utility."
3. The human retains decisional authority — they make "the final choice" from the recommendation set.

All three assumptions are transitional artifacts that collapse under the Discontinuity Thesis.

Under P1 (Cognitive Automation Dominance), the human is not the initiator. The AI initiates. The human's "preference message" is itself generated, curated, or subsidized by AI systems that already know what the human wants better than the human does. The "costly and noisy message" is a human-facing fiction maintained for interface legitimacy. In reality, the AI's posterior belief will be formed from behavioral data, inferred preferences, and model priors — the "communication" is theater.

Under P2 (Coordination Impossibility), the assumption that institutional or contractual arrangements can preserve the "user makes the final choice" framing at scale is fiction. Once AI-driven recommendation systems mediate the vast majority of consumption decisions — as they are already doing — the human's "choice" from a set becomes a performed veto, not a substantive decision. The paper optimizes this performance.

HIDDEN ASSUMPTIONS

• Human preference is stable and exogenous. The paper assumes the user has a "true preference" in d-dimensional space that the AI must infer. This is pre-DT preference theory — it treats preferences as fixed attributes rather than dynamically constructed through AI-mediated exposure.
• Utility maximization is the operative frame. The entire analysis is Pareto-optimal under welfare economics. This framework is structurally unable to model the distribution of viability — the difference between a Sovereign who benefits from this system and a Servitor who is merely consuming optimized recommendations.
• The product universe is fixed. The paper samples from a "product universe" without modeling what happens when AI systems generate the product universe itself — when recommendations are synthetic goods, services, and experiences that did not exist before the AI generated them. This is already happening in code, art, music, and will expand into material production.
• The user bears a "search cost." Implicitly, the user is doing productive economic work in the search process. This is the last fragile thread connecting the paper to any productive economy. The DT thesis predicts this cost goes to zero — AI eliminates search costs entirely — which maximizes the efficiency of the recommendation system while eliminating the human's last structural role in the transaction.

SOCIAL FUNCTION

Prestige signaling within the academic AI community — the paper performs technical rigor on a problem that is already being solved by systems deployed in production, without the Bayesian overhead. It also serves as transition management material: it reassures that the human is still in the loop, still making choices, still the principal. The "final choice" framing is ideological anesthetic dressed in mutual information cost functions.

The authors are competent engineers doing legitimate mechanism design work. The social function of the paper, however, is to make the architecture of human displacement feel like a well-optimized welfare problem.

THE VERDICT

Mechanical Death: The paper's optimization target — maximizing user utility given communication and search costs — is itself a cost function for the last mile of human economic participation. When AI eliminates search costs to near-zero (which is the direction of every deployed system), and when AI infers preferences from behavioral data rather than "costly messages," the user's utility-maximizing strategy converges on: accept whatever is recommended. The "final choice" becomes vestigial.

Social Death: The paper's framework requires humans to have preferences worth inferring and utility to maximize. Under the DT, the mass of consumers become attention sellers and recommendation performers, not economic principals. The paper's most efficient policy — "optimal interaction policy uses only one of communication and search" — is a mathematical description of the Hyena's Gambit: pick the cheapest path to consumption, which under AI systems is to be a recipient, not an initiator.

The paper is technically sound mechanism design for a system whose terminal function is the dissolution of the economic role it assumes as its premise.