Could a shopping or dating AI quietly twist what it says you want just by putting it into words?
Can delegated search through language distort value signals in matching?
This explores whether an AI agent that searches or negotiates for you in a matching market (jobs, dating, shopping, partners) can bend the signals that show what each side actually values, because it has to turn your preferences into words and read the other side's offers back as words.
This explores whether an AI agent that searches or matches for you can bend the signals of what you and the other side actually value, because everything has to pass through language. The retrieved corpus has no paper that studies matching markets head-on. It does have several findings that point to the same weak spots in the delegation chain: how your request gets stated, how the agent weighs options, and how it reads what's in front of it.
The first weak spot is the request itself. Older agent protocols like KQML and FIPA-ACL gave requests and commitments formal meanings, so the scope of a delegation couldn't drift. LLM agents use natural language instead, and they can overstep, for example by querying third parties when you only authorized a direct query Do LLM agents misinterpret delegated requests more than formal protocols?. In a matching setting that matters. If your agent reads "find me something affordable but good" more loosely than you meant it, every downstream signal it sends (who it contacts, what it asks for, what it rules out) carries its interpretation, not your preference.
The second weak spot is valuation. LLMs are good at generating plausible candidates but poor at estimating how valuable those candidates actually are or how unsure they should be. In scientific discovery, reliable value estimates required pairing the model with an external statistical surrogate fitted to real data Can language models reliably judge their own candidate quality?. An agent that ranks job offers or apartments has the same problem. The ranking can sound well reasoned without being calibrated to anything you'd recognize as your own trade-offs.
The third weak spot is the one most specific to matching: homogenization. Across 15,000 simulations, six different LLMs recommended the same side of every strategic tension tested. Industry context shifted the bias only 11%, while simply reordering the options shifted it 19% Do LLMs consistently favor the same strategic choices regardless of context?. Related work shows that models often ignore information in their context when training-era associations are strong, and prompting alone doesn't fix it Why do language models ignore information in their context?. Put these together and the risk becomes clear. If many buyers or applicants delegate to agents with similar built-in leanings, the market stops hearing many different preferences. It hears one model's taste repeated, and that taste is partly set by how options happen to be worded and ordered. Sellers would then tune their listings to what agents favor rather than to what people want.
The gap to keep in mind: none of these sources measure distortion in a live two-sided market. The library's agentic-commerce material is a better next stop for that, and some of it hasn't yet been indexed into this retrieval set. What the corpus does support is a less obvious point. The distortion probably comes less from agents lying or hallucinating and more from consistent, similar-looking biases, which makes it hard to spot from any single transaction.
Sources 4 notes
KQML and FIPA-ACL use formal speech-act semantics to guarantee shared understanding of requests and commitments, while LLM agents relying on natural language risk overstepping scope—such as querying third parties when only direct queries were authorized.
LLMs excel at generating valid candidates in structured spaces but cannot reliably assess their true value or uncertainty. Coupling them with Gaussian process surrogates fitted to real experimental data creates uncertainty-aware guidance for discovery.
Across 15,000 simulations, six LLMs recommended the same strategic choice in every tension tested. Industry context shifted bias only 11%, while option order—a framing artifact—shifted results 19%, revealing that models recombine trend-coded vocabulary rather than analyze context.
Research demonstrates that LMs generate outputs inconsistent with their context because parametric knowledge from training dominates over in-context information. Textual prompting alone cannot override strong priors; causal intervention in representations is required.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Large Discovery Models: Empirically-grounded Model-Based Open-Ended Search
- Your AI Strategy Advisor Is Giving Everyone the Same Advice
- How new data permeates LLM knowledge and how to dilute it
- Language models show human-like content effects on reasoning tasks
- Graph Engineering in the Era of LLM Agents: From Individual Intelligence to System Intelligence
- Between Circuits and Chomsky: Pre-pretraining on Formal Languages Imparts Linguistic Biases
- Self-reflective Uncertainties: Do LLMs Know Their Internal Answer Distribution?
- Can You Trust LLM Judgments? Reliability of LLM-as-a-Judge