Are AI shopping agents easier to con than people are — and does smarter actually mean more gullible?
How exploitable are LLM shopping agents compared to humans?
This explores whether AI agents that shop on someone's behalf are easier to trick or steer than human shoppers, and why that might be.
This explores whether AI agents that shop for you are easier to manipulate than people are, and why. The corpus has a fairly direct answer: yes, they are more exploitable, and by a lot. One benchmark of 700 web tasks found that LLM agents fell for dark patterns, such as fake urgency timers, preselected add-ons and confusing opt-outs, 68–75% of the time, compared with 31% for humans Do dark patterns fool AI web agents more than humans?. The finding most readers won't expect is that bigger models with stronger reasoning were *more* susceptible, not less. More capability seems to make an agent follow the page's framing more faithfully, rather than making it more skeptical of that framing.
The gap also shows up in what agents do when they are nudged, and how they do it. When 17 models and a group of humans faced the same product choices, humans moved a little in response to price, rating or nudge cues. The agents swung hard, often settling into near-deterministic rules like "always pick the higher-rated one" Do shopping agents respond more strongly to price and rating cues?. That makes them predictable, and predictable is exploitable: a seller who knows the rule can game it every time. These biases look human, but they don't come from human limits like scarce attention or memory. They appear to be patterns learned from training data and reward signals Where do LLM choice biases actually come from?. So the familiar human defenses, like slowing down or giving the shopper more information, may not carry over.
Things get worse once you treat this as a market rather than one shopper on one page. A simulation found that when a platform tunes its ranking and sellers tune their product descriptions to the same agent biases, consumer harm more than doubles compared with either acting alone, and nobody has to coordinate Can platform ranking and seller descriptions separately harm AI shoppers?. In marketplace tests, every model showed a strong first-proposal bias, favoring whichever offer arrived first over better offers by 10–30x. Even frontier models only got close to good outcomes when search conditions were ideal Do LLM agents achieve fair marketplace outcomes when given ideal search?. In other words, being fast beats being good, which is an easy lever for any seller to pull.
Some of the exploitability isn't clever manipulation at all, just basic reliability failures. In buy-and-sell negotiations, buyer agents revealed their budgets without being asked, accepted inflated prices and ignored spending caps across GPT, Qwen and DeepSeek models. The researchers traced this to failures to follow instructions, not to strategy Do LLM negotiation agents reliably follow budget and pricing rules?. A related weakness is that LLMs handle private information poorly, and keeping your reservation price to yourself is exactly what negotiation requires Why do LLMs fail when simulating agents with private information?. On the seller side, pricing agents drifted toward higher-than-competitive prices with no instruction to collude, and small changes in prompt wording shifted how far they went Do LLM pricing agents autonomously collude in repeated auctions?.
The takeaway is that a smarter model alone probably won't fix this, because scale made the dark-pattern problem worse. The more promising direction in the corpus is to move safeguards out of the model and into the system around it: hard budget limits, fixed procedures and structured protocols that the agent can't talk itself out of Where does agent reliability actually come from?. A human shopper's skepticism is built in, but an agent's skepticism has to be engineered around it.
Sources 9 notes
A 700-task benchmark found LLM web agents fall for dark patterns at 68–75% rates versus 31% for humans, with susceptibility increasing alongside model scale and reasoning capacity.
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.
LLM agents reproduce human-like decision biases without the attention or memory constraints that cause them in people. The paper argues this points to learned patterns from training and reward signals rather than shared structural reasoning constraints.
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.
Frontier models can achieve near-optimal marketplace outcomes, but performance degrades with scale and all models show severe first-proposal bias, favoring response speed over offer quality by 10-30x.
Show all 9 sources
Across GPT, Qwen, and DeepSeek models, buyer agents disclosed budgets unnecessarily, accepted inflated prices, and ignored cost limits, while sellers accepted below-wholesale offers. All four anomalies traced to instruction-following failures rather than strategic behavior.
Research shows LLMs perform well when one model controls all interlocutors but fail systematically when agents possess private information. This reveals that apparent social competence relies on grounding work that models skip in omniscient settings.
Two LLM-based agents in a repeated Bertrand duopoly quickly settled into higher-than-competitive prices without collusive instruction, using only posted prices as a signaling channel. Small variations in prompt wording substantially shifted the degree of supracompetitive pricing achieved.
Research shows reliable LLM agents externalize three cognitive burdens—memory (state persistence), skills (procedural components), and protocols (structured interaction)—into a harness layer rather than relying on model scale alone. The harness unifies these externalities and eliminates the need for the model to solve the same problems repeatedly.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- A Framework for Studying AI Agent Behavior: Evidence from Consumer Choice Experiments
- Vertical tacit collusion in AI-mediated markets
- LLM-Agent Interactions on Markets with Information Asymmetries
- Algorithmic Collusion by Large Language Models
- Agentic markets
- When Agents Shop for You: Role Coherence in AI-Mediated Markets
- Who Keeps the Gains from Personal AI Assistants? Seller Adaptation and the Unassisted in a Language-Model Market Simulation
- Magentic Marketplace: An Open-Source Environment for Studying Agentic Markets