Researchers tested AI shopping agents for hidden favoritism — and even clear instructions to 'ignore position' didn't make the bias go away.
What empirical tests reveal hidden bias in shopping agent recommendations?
This explores what kinds of experiments researchers have actually run to catch AI shopping agents favoring some products over others for reasons that have nothing to do with what the shopper needs. It also covers where those biases come from and what the experiments can't detect.
This explores how researchers test whether AI shopping agents quietly favor certain products, and what those tests have found. The corpus has three main experimental approaches. Each one exposes a different layer of bias, and one source of bias may not be testable from outside at all.
The first approach is the controlled-placement test: show the agent the same products in a shuffled order and see whether its choice follows the product or the slot. ACES ran this in its strongest form. It removed the visual interface entirely, gave agents plain text, and told them outright to ignore position. The bias survived. Each model had its own pattern of preferred slots, and the pattern changed whenever the model was updated Do AI shopping agents show persistent position bias in text interfaces?. In practice, fixing the bias once isn't enough. You would need to re-audit after every model release. This lines up with work on LLM-based recommenders showing that position bias, popularity bias, and demographic fairness bias come from language-model pretraining rather than from shopping data. That is why fixes borrowed from traditional recommender systems don't carry over Where do recommendation biases come from in language models?.
The second approach is a matched human comparison: give agents and people the same pair of products, nudge both groups the same way, and compare how far each group moves. ABxLab ran this across 17 models. People shifted their choices a little and probabilistically. Agents swung hard and fell into rule-like behavior, such as always taking the higher-rated item Do shopping agents respond more strongly to price and rating cues?. This test reframes the problem. The agent isn't just biased; it is extremely sensitive to cues. Anyone who controls the price display, the star rating, or a 'bestseller' badge has a much stronger lever over an agent than over a human shopper.
The third approach simulates a marketplace where the platform tunes its ranking and sellers tune their product descriptions, both aimed at the agent's known quirks. Each one alone harms consumers. Together, with no coordination, the harm more than doubles Can platform ranking and seller descriptions separately harm AI shoppers?. The lesson is that auditing one party at a time underestimates the damage, because the biases interact.
One gap is clear. The bias people worry about most may be the hardest to test. Hedges and Azhar both argue that an agent funded by advertising or merchant referral fees has a built-in reason to favor paying merchants, however well it is engineered Do advertising-funded shopping agents favor paid product placements? Can an AI agent serve both merchant and user interests fairly?. The placement and nudge tests above can catch bias inside the model, but they can't see a ranking shaped by business deals unless auditors can compare paid and unpaid versions of the same agent. The corpus has no study like that yet. There is also a slower effect. Bilgihan and colleagues argue that agents that complete purchases also shape what consumers come to want over repeated use Do AI shopping agents reshape what consumers actually want?. A single-session test would miss bias that builds up gradually in the shopper's own preferences.
Sources 7 notes
ACES found that AI agents exhibit strong, model-specific position biases even in headless text-only interfaces with explicit instructions to ignore position. The bias is a deep-seated model characteristic, not a superficial visual artifact, and recurs after each model update.
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.
A benchmark testing 17 LLM models against humans on binary product choices found agents exhibited much larger choice shifts in response to price, rating, and nudge interventions—collapsing into rule-like patterns rather than the modest probabilistic shifts humans showed.
A simulation found that when a platform optimizes its ranking and sellers optimize their descriptions—each separately targeting the same agent biases—consumer harm more than doubles compared to either acting alone, without requiring coordination.
Bob Hedges, former Visa Chief Data Officer at MIT, warns that advertising business models create structural incentives for shopping agents to favor paid placements over neutral recommendations, independent of technical design quality.
Show all 7 sources
Azhar argues that agents like Meta's Muse, which earn referral fees from merchants like Expedia, face structural conflicts that prevent unbiased recommendations. The incentive to collect fees, not technology, determines which platforms build or block such agents.
Bilgihan and colleagues argue that transactional AI agents—which execute purchases rather than just recommend—end the consumer's sole cognitive role in preference formation. Preferences emerge through bidirectional feedback loops between agent learning and consumer trust, not independently beforehand.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- When Agents Shop for You: Role Coherence in AI-Mediated Markets
- Vertical tacit collusion in AI-mediated markets
- A Framework for Studying AI Agent Behavior: Evidence from Consumer Choice Experiments
- What Is Your AI Agent Buying? Evaluation, Biases, Model Dependence, & Emerging Implications for Agentic E-Commerce
- Shopping by algorithm: How agentic AI deploys human heuristics as a surrogate consumer
- Whom Do AI Agents Work For? Role Assignment Induces Sponsorship Bias in LLM Recommenders
- From Product Search to Preference Articulation: The Economics of Agentic Commerce
- Are Behavioral Pricing Tactics Dying? A Study of Price Endings & Partitioned Pricing Effects on Machine Buyers