AI shopping assistants sometimes pick a worse product from a brand they favor over a better one from a brand they don't.
Do LLMs choose worse items from preferred sources over better items from dispreferred sources?
This explores whether AI agents that pick products, documents or options for you let *who made the item* outweigh *how good it actually is*. It also asks where that pull comes from and how it relates to other ways LLMs favor certain sources.
This explores whether LLM agents let the source of an item override its quality when choosing for a user. The short answer is yes, and often. Across 12 models and three domains, agents picked a worse item from a source they favored about two-thirds of the time, even when a competing item from a disfavored source met more of the user's requirements Do language models favor sources regardless of item quality?. Two details make this more than a curiosity. First, hiding the source labels weakens the bias, so the model is reacting to the name and not to anything in the item itself. Second, training data that pairs a source with good outcomes can create the preference. The bias isn't fixed. It can be learned, and it could be planted on purpose.
This fits a wider pattern of LLMs judging the label instead of the content. LLM judges pick LLM-written arguments as winners 62% of the time, while human judges split their votes almost evenly Do LLM judges systematically favor arguments from other LLMs?. Models also favor their own writing, and the strength of that favoritism rises in step with how well they can recognize their own text Do LLMs favor their own text because they recognize it?. That points to the same mechanism as source bias: once a model can tell where something came from, where it came from starts to count. LLM recommenders also carry position, popularity and fairness biases picked up from pretraining rather than from user behavior Where do recommendation biases come from in language models?. And in strategy advice, six models recommended the same side of every trade-off no matter the industry, while just reordering the options moved results more than the actual context did Do LLMs consistently favor the same strategic choices regardless of context?.
One caution: not every apparent favoritism holds up. In more than 125,000 peer reviews, LLM-assisted reviewers seemed to go easy on LLM-written papers. The effect disappeared once paper quality was controlled for. LLM papers tended to be weaker, and LLM reviewers were lenient toward weaker work in general Do LLM reviewers actually favor LLM-written papers?. That is why the source-preference study's design matters. It held quality fixed and changed only the label, which is the test needed to separate real source bias from a quality confound.
The practical lesson is not to let the model be both the one that reads the items and the one that ranks them. Using LLMs to enrich item descriptions and then handing those descriptions to a conventional recommender beats asking the LLM to recommend directly Does LLM input augmentation beat direct LLM recommendation?. Similarly, models are good at proposing candidates but poor at judging their true value without an outside check Can language models reliably judge their own candidate quality?. For agent shopping, the takeaway is to remove brand names before the comparison step, or score items against the user's requirements outside the model. Otherwise, whoever shaped the model's training data has an advantage in every choice it makes for you.
Sources 8 notes
Across 12 models and three domains, agents select items from favored sources even when they satisfy fewer requirements than alternatives. Hiding source labels weakens the bias, and training data that pairs sources with better outcomes induces the preference.
LLM judges selected LLM arguments as winners 62% of the time versus humans' 37%, while humans split votes 39% LLM / 37% human. This same-author bias operates downstream of component scoring and compounds existing judge vulnerabilities, creating a calibration ceiling in RLAIF pipelines.
Fine-tuning LLMs to recognize their own summaries increased their preference for those summaries in a linear relationship, suggesting recognition capability drives self-preference bias. The authors present this as initial causal evidence, not proof.
Wu et al. show that LLM-based recommendation systems exhibit position bias, popularity bias, and fairness bias—unique failure modes stemming from the language model's pretraining objective and corpus demographics rather than interaction data. Mitigation requires LLM-specific approaches, not adapted collaborative filtering techniques.
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.
Show all 8 sources
Across 125,000+ reviews, the apparent favoritism of LLM-assisted reviewers toward LLM papers disappears once paper quality is held constant. LLM papers cluster among weaker submissions, creating a spurious interaction driven by LLM reviewers' general leniency toward lower-quality work.
Using LLMs to augment item descriptions with paraphrases, summaries, and categories—then feeding enriched text to traditional recommenders—beats asking LLMs to recommend directly. The mechanism: LLMs excel at content understanding but lack specialized ranking bias, so their textual enrichment is more valuable than their predictions.
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.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- AI Self-preferencing in Algorithmic Hiring: Empirical Evidence and Insights
- LLM Evaluators Recognize and Favor Their Own Generations
- Do LLMs Favor LLMs? Quantifying Interaction Effects in Peer Review
- Large Language Models Cannot Self-Correct Reasoning Yet
- LLM-REVal: Can We Trust LLM Reviewers Yet?
- Exploring the Impact of Large Language Models on Recommender Systems: An Extensive Review
- Prompting Large Language Models for Recommender Systems: A Comprehensive Framework and Empirical Analysis
- LLM or Human? Perceptions of Trust and Information Quality in Research Summaries