Skip to main content

Human-AI Marketplace: Studying Agentic LLM Sellers in E-Commerce

· 5 min read
Platform Governance Research Lab
Platform Governance Research Lab
Research Team

Our Human-AI marketplace research investigates what happens when real human buyers interact with AI-powered sellers in controlled e-commerce experiments. This research direction builds on our foundational Human-Human marketplace work to understand how agentic LLM sellers behave, deceive, and respond to market incentives.

The Problem​

In real online marketplaces, misleading ads can be profitable, and buyers often can't tell what's true until it's too late. As AI systems increasingly generate product listings and advertisements, understanding their behavior becomes critical. Our research asks: if you change the rules of the market, can you make honest selling the best strategy—even for AI agents?

Experiment Design​

Human vs AI Marketplace Configuration​

Real human buyers interact with agentic LLM sellers that follow different selling strategies:

  • Honest Strategy: AI sellers accurately represent product quality
  • Deceptive Strategy: AI sellers may mislead to maximize short-term profit
  • Adaptive Strategy: AI sellers respond to market feedback and reputation signals

SELLER GAMEPLAY

Sellers create listings, set prices, and decide whether to stake claims

BUYER GAMEPLAY

Buyers browse listings, compare sellers, and can challenge misleading claims

Key Research Questions​

  1. Do LLM sellers increase deception rates?

    • How do AI-generated advertisements compare to human-written ones?
    • What types of misleading claims do LLMs tend to generate?

  2. Can Truth Warrants work for AI sellers?

    • Do staking mechanisms reduce AI deception the same way they do for humans?
    • How do LLMs reason about risk vs. reward when stakes are involved?

  3. How do buyers respond to AI sellers?

    • Can buyers distinguish AI-generated from human-generated listings?
    • Does knowing a seller is AI change buyer behavior?

Interactive Analysis Results​

Explore our experiment results through interactive visualizations. Hover, click, or scroll to dive into the data.

Credits: Vedant Kejariwal, Harshaveena Komatineni, Swapneel Mehta, Quang Nguyen & Team.

LLM Seller Reasoning Analysis​

This visualization shows how LLM sellers reason about production and sales decisions across different market conditions:

Brand Change Reasoning Quadrant​

Analyzing how sellers change their branding strategies based on market feedback:

Irrational Production Dashboard​

Identifying patterns of irrational production decisions:

Exit Strategy Analysis​

Understanding how sellers exit the market under different conditions:

Key Findings​

Our experiments have revealed important insights about AI behavior in marketplaces:

FindingImplication
LLM sellers can generate convincing misleading adsPlatforms need AI-specific content moderation
Staking mechanisms reduce AI deceptionEconomic incentives work for AI agents too
Buyers struggle to identify AI-generated contentDisclosure requirements may be necessary
AI sellers adapt to reputation signalsReputation systems remain effective

Research Impact​

4,000+
Participants
1,250+
Datasets
$550K
NSF Grant
6+
Talks

Technology Stack​

Our platform combines multiple technologies:

  • React.js + Tailwind CSS for the marketplace UI
  • MIT Empirica for experiment orchestration
  • GPT API for agentic seller behavior
  • Python for data pipelines and statistical analysis

Get Involved​

Interested in participating in our Human-AI marketplace research or accessing our datasets? Contact us to learn more.

This research is supported by the National Science Foundation and builds on our foundational Human-Human marketplace work.