Following the Preference, Missing the Optimum: Compliance Without Optimization in AI Housing Recommendation

TEXT START: Large language models are becoming the first point of contact for consumer search in domains where the stakes are material and the law is explicit.

The Dissection

This is an empirical separation of two capabilities that AI vendors prefer to blur: obeying a stated preference and actually optimizing a choice.

The study constructs a constrained ground truth from 150 synthetic New York renter scenarios and 120 real listings per scenario. Its result is decisive: constraint violations are rare, but 39.0% of recommendations are strictly dominated by another listing in the same pool. The rejected alternative is typically $900 cheaper and 3.5 minutes faster. Changing the user’s preference changes the recommendation in the expected direction, yet the model still selects options substantially above the cheapest qualifying listings. Even explicit lexicographic instructions do not repair the failure.

The system listens. It does not reliably search, compare, or optimize. That is the paper’s real contribution.

The Core Fallacy

The fallacy is a category error, not necessarily bad measurement. The paper frames a structural allocation problem as a ranking defect. Dominance-rate instrumentation can expose avoidable waste, but a better ranker merely makes automated gatekeeping more efficient. It does not address who owns the inventory, the recommendation channel, the data, or the surplus created by controlling access.

Under the Discontinuity Thesis, even a perfectly optimizing recommender would not preserve productive participation. It would simply become a more powerful intermediary between people and essential goods.

There is also a narrower methodological limit: “optimum” is too large a word for rent, bedrooms, and commute time. Strict dominance proves inferiority on measured dimensions, not global inferiority once safety, condition, neighborhood, fees, landlord behavior, availability, and unobserved preferences enter the frame. The strongest claim is avoidable measured inferiority—not discovery of the objectively best home.

Hidden Assumptions

• The candidate pool is representative, stable, and equally available to the renter.
• Rent, bedrooms, and GTFS commute capture enough of housing utility to judge meaningful optimization.
• A cheaper, faster, no-smaller listing is genuinely preferable despite unmeasured attributes.
• Synthetic renter scenarios adequately represent real behavior, constraints, and legal exposure.
• The model is the primary source of failure, rather than inventory quality, interface design, paid placement, retrieval systems, or landlord incentives.
• Vendors can deploy optimization without strategic manipulation, commercial capture, or new forms of steering.
• A user’s stated preference is an adequate proxy for welfare.
• Improving recommendation quality solves the material problem instead of merely making users more dependent on the recommendation layer.

Social Function

Classification: partial truth functioning as transition management.

The paper is not copium. It documents a real and costly failure with unusually clean instrumentation. But it converts a question of power into a quality-assurance metric. “Who should control access to housing search?” becomes “How do we reduce dominance rate?” That translation is useful to deployers because it offers a repairable defect while leaving ownership, concentration, and dependency untouched.

It is an audit of the machine’s incompetence, not an audit of the regime that installs the machine as gatekeeper.

The Verdict

This is a sharp, narrow indictment. AI housing recommenders already display the dangerous combination of authority without competence: they can respond to preferences well enough to earn trust while systematically failing to identify better available options.

The result implies avoidable rent exposure and lost opportunity for users. It does not imply that optimization will restore autonomy, democratize housing access, or reverse the post-WWII economic break. Dominance-rate instrumentation can make the machine less stupid. It cannot make the mass employment–wage–consumption circuit live again.