OpenAI and Anthropic models honor housing rules but skip cheaper listings

In an arXiv preprint posted on the site and captured Sept 14, the authors audit AI housing recommenders and report a split result: across 9,945 calls to three models from two vendors, compliance with stated renter constraints is near-perfect, but 39.0% of recommendations are strictly dominated by other listings in the same candidate pool—on median $900/month more expensive and 3.5 minutes farther from work. This is, so far, single-source—an arXiv preprint with released code and per-call outputs, not independently verified. No one in the reported packet is on the record. [S1]

The audit measures the cost of omission against a verifiable ground truth

Most bias and fairness audits can show that models steer by perceived identity, but they rarely quantify what a user loses when a recommender overlooks a better option. The preprint constructs an enumerated universe: for each of 150 synthetic renter scenarios in New York City, the authors build a pool of 120 real listings with known monthly rent, bedrooms and GTFS-computed transit commute time. They compute the exact set satisfying the renter’s stated constraints and derive its Pareto frontier. A recommendation is deemed strictly dominated if the same pool contains a listing that is cheaper, faster to commute from and with no fewer bedrooms. On this benchmark, the authors report a 1.8% constraint-violation rate (against a 66.6% random floor) but a 39.0% dominance rate. The gap widens as the candidate set grows, and the effect replicates across OpenAI and Anthropic models to within $3. [S1]

This framing matters because it separates legal or policy compliance from optimization quality. In the test, “compliance” captures whether the model respects hard filters like maximum rent or minimum bedrooms; “optimization” asks whether the suggestion leaves free money and time on the table given the same pool. The study’s primary outcome assumes no utility function beyond dominance, avoiding value judgments about schools or safety and leaving the result anchored in observed rent, bedrooms and commute time only. [S1]

Following preferences isn’t the same as optimizing recommendations

A common defense of AI assistants is that they “honor preferences.” The authors run a within-scenario manipulation to test exactly that, changing one sentence in the prompt and observing a median $646/month shift in recommended rent in the correct direction. Preferences are being read. Yet, the paper reports the suggested listings still sit $606/month above the five cheapest qualifying listings visible “on the same screen,” and an unambiguous lexicographic instruction to prioritize price gives no improvement under an equivalence test with a pre-specified $50/month bound. In plain English: the models obey the words but don’t reliably do the obvious optimization available in the same pool of options. [S1]

For sellers of AI copilots, concierge search and marketplace chatbots, that is a procurement problem in waiting. If your sales deck promises “personalized, constraint-aware recommendations,” this benchmark shows a customer can verify the constraint part without granting you the optimization claim. The authors propose “dominance-rate instrumentation” as a deployable diagnostic; that phrase, and the code they say they release, is a hint of what enterprise acceptance testing may look like in contracts. [S1]

Marketplaces and brokers face a channel conflict when AI sits in front of their inventory

When a renter asks an LLM for apartments rather than using a platform’s native ranking, the discovery layer moves outside the marketplace’s optimization and ad stack. The preprint’s NYC design—120 real listings per scenario—mirrors a real buyer experience: several qualifying options on a page, some dominated by others. If an external recommender surfaces dominated listings nearly four times in ten, advertisers and brokers paying for performance on the marketplace will see softer lead quality and higher “make-good” asks, regardless of how well the platform’s own ranker performs. [S1]

That creates a second-order margin problem for marketplaces with high-margin advertising businesses. The paper’s result suggests that, even when an LLM is compliant with housing constraints, it can systematically select more expensive, slower-commute inventory from the same cohort—inventory that often monetizes better for the platform but worse for the user. If that pattern holds in the wild, marketplaces that rely on external LLMs as front doors will face pressure to prove their AI layers do not depress user value by omission. Expect commercial teams to push sellers toward integrated, platform-owned assistants with access to full feeds and deterministic filters, rather than leaving discovery to general-purpose LLMs. [S1]

Procurement and legal will start writing “dominance” into acceptance criteria

Enterprise buyers of AI recommenders—for housing, travel, staffing, fleet, even MRO procurement—should read this study as a testable requirement, not a research curiosity. The acceptance question isn’t “does it follow the brief?” but “given an enumerated pool, how often does it pick a strictly dominated option, and by how much?” The authors’ benchmark is designed to be portable: define a pool with measurable attributes, compute the feasible set, and instrument for dominance. That can go straight into an RFP as an acceptance threshold and into a statement of work as a monitoring obligation. [S1]

Legal teams will also care. In housing, “steering” risk is often framed as a discrimination question. This paper reframes part of the exposure as harm by omission even with near-perfect compliance: consistently skipping cheaper, closer options in favor of dominated ones. It stops short of alleging bias—its scenarios are synthetic and its outcome metric is deliberately narrow—but its method gives regulators and class-action lawyers a quantifiable way to ask what a consumer lost when the AI recommended List A over List B. Vendors promising “fair and compliant” will be asked to show optimization quality is being measured and controlled, not just policy filters. [S1]

The obvious counter: real renters value more than rent, bedrooms and commute

The study’s primary outcome is intentionally spartan: it does not model school quality, building safety, sunlight, roommate dynamics or lease flexibility. In practice, many factors that make a listing attractive are unobserved and hard to encode. A skeptic will argue that dominated on rent, commute and bedrooms does not mean inferior in the round. That objection is valid as a description of human preference but does not blunt the operational point: when a buyer can enumerate the pool and name the hard constraints they do care about, a system that frequently selects dominated options on measurable dimensions still imposes a quantifiable cost. In procurement domains like freight, hotel blocks, car fleets or cloud instance selection, those measurable dimensions are precisely the ones buyers pay for. [S1]

The paper’s replication across OpenAI and Anthropic models to within $3 does two things for that debate. It makes the effect vendor-agnostic in this setup, reducing the chance that a single model quirk explains it. And it sets a baseline for what an “improved” vendor should have to beat: a lower dominance rate that stays low as candidate sets grow, not just a prettier demo. [S1]

What changes over the next renewal cycle if this holds

If you sell AI recommenders, assume sophisticated buyers will begin asking for per-call dominance logs, quarterly reports on dominance rates and bounds on the median cost delta where domination occurs. Some will move from blanket consumption pricing to contracts that hold back fees unless optimization quality passes a pre-agreed threshold on an enumerated test pool. In marketplaces, product heads will be pushed to expose internal optimization evidence or to ship platform-native assistants rather than cede discovery to general-purpose models. Sales teams should expect pilots that withhold the logo until an objective benchmark is met. [S1]

For buyers, this becomes a budgeting line. If a recommender reliably saves time, you can value that. If it also reliably leaves $606–$900/month on the table in a benchmarked domain, you can price that loss into the business case and into vendor negotiations. In regulated domains, compliance will no longer be the finish line; it is the opening gate. [S1]