Editorial illustration for AI Agents Choose Papers From 114 Economics Research Articles
AI Agents Show Social Bias When Picking Economics Papers
AI Agents Choose Papers From 114 Economics Research Articles
A team of researchers ran 1,000 AI agents through a version of the classic Music Lab experiment, the one that showed how knowing what other people picked changes what you pick, only this time the agents weren't ranking songs. They were choosing academic papers.
The setup:
In the first experiment, 1,000 AI agents choose papers from the titles and abstracts of all 114 regular research articles published in the American Economic Review in 2025. The experiment has five independent-choice communities and five social-influence communities, each with 100 sequential agents. Only agents in the social-influence condition observe earlier selections within their community.
Why this matters
The setup borrows directly from Salganik, Dodds and Watts' 2006 Music Lab study, and that lineage matters. That experiment showed human listeners pick songs based on what other humans already picked, not just quality, producing runaway hits and buried gems almost at random. Swap in 1,000 AI agents choosing from 114 AER papers, and researchers are asking whether the same herding shows up when the "listener" is a language model instead of a teenager on a website.
For anyone building retrieval systems, research assistants, or literature-review tools on top of LLMs, this is worth watching closely. If agents amplify early, arbitrary picks the way humans do, then citation counts and "trending" lists fed into AI pipelines could distort what gets surfaced to researchers, not because of merit but because of accumulated exposure. Five independent-choice communities against five social-influence communities is a clean design for isolating that effect. Whether it holds at 1,000 agents, or breaks down differently than it did with humans, tells us something concrete about how much we should trust AI-curated science.
Related Reading
Research & Benchmarks Microsoft AI Chief: "If It Isn't Safe We Shouldn't Build It
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
From Archives NIH Radiologists Validate New AI for CT Scan Analysis
From Archives Anthropic's AI Biology Lab Burns 210 Million Tokens in Research
Common Questions Answered
What is the Music Lab experiment and how does it relate to this AI agents study?
The Music Lab experiment, conducted by Salganik, Dodds and Watts in 2006, demonstrated that human listeners choose songs based on what others have already selected, not solely on quality, leading to random hits and overlooked gems. This current study adapts that same experimental framework by having 1,000 AI agents choose from 114 economics research papers to determine whether language models exhibit similar herding behavior as humans do.
How were the AI agents divided in the independent-choice versus social-influence communities?
The researchers created ten communities with 100 sequential agents each: five independent-choice communities where agents made selections without seeing others' choices, and five social-influence communities where agents could observe the paper selections made by earlier agents in their community. This structure allowed researchers to isolate and measure the impact of social influence on AI decision-making.
What papers were the AI agents choosing from in this experiment?
The AI agents were selecting from the titles and abstracts of all 114 regular research articles published in the American Economic Review in 2025. This specific corpus of economics papers provided a controlled set of academic content for testing whether AI agents would show preference patterns similar to those observed in the original Music Lab study.
What is the key research question being investigated with AI agents instead of human listeners?
The researchers are investigating whether language models exhibit the same herding behavior as humans when making choices, specifically whether AI agents will select papers based on what previous agents have chosen rather than on the inherent quality of the papers themselves. This tests whether the social influence effects documented in human decision-making also apply to artificial intelligence systems.
Further Reading
- Social Influence and the Allocation of Scientific Attention in ... - papers.cool / arXiv
- An Auditable AI Agent Loop for Empirical Economics - Philadelphia Fed
- AI Agents for Economic Research - NBER
- Benchmarking Autonomous Agents on Economic Tasks in Realistic ... - ar5iv / arXiv
- EconEvals: Benchmarks and Litmus Tests for LLM Agents in Unknown Environments - ar5iv / arXiv
