When AI agents saw what was popular, early choices overwhelmed their own judgement
In simulated music markets involving 168,480 LLM agents, showing cumulative popularity made outcomes less predictable and more unequal across all 13 models. It is a controlled preprint, not evidence about a live platform.
The short version
- What happened
- Researchers ran 13 language models through 17 simulated markets each: one independent world of 1,440 agents and 16 socially influenced worlds of 720 agents, totalling 168,480 agent decisions.
- Who it affects
- People may experience a system's apparent consensus as evidence of quality even when it began as an arbitrary early lead. That matters wherever agent groups filter culture, products, applications or public information.
- What remains unproven
- The paper is an unreviewed preprint and reports a controlled simulation rather than a deployed platform outcome.
Keep in mind
The paper is an unreviewed preprint and reports a controlled simulation rather than a deployed platform outcome.
All agents within a world used the same underlying model; real systems can mix models, humans and changing objectives.
Agents saw cumulative counts but did not discuss choices, share memory or face social or economic incentives.
The human comparison used different media and procedures, so it is contextual rather than a like-for-like performance test.
Source analysis · 1 linked source. Primary analysed source date: 22 October 2026.
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Read the full report · 4 min The evidence, methods, implications and limitations
Popularity changed which songs won
The researchers adapted the classic Music Lab experiment for language models. Each agent received a persona and a fixed budget of four selections from 48 unfamiliar, generated song lyrics. For every model, one independent world hid popularity. Eight weak-influence worlds displayed cumulative recommendation counts in random order, while eight strong-influence worlds also ranked songs by those counts. Across the 13 models, that produced 12,960 agents per model and 168,480 agents overall.
Social influence increased the variation between repeated worlds for every model at both influence levels, with reported p-values below .001. In practical terms, the same songs and model could produce different winners depending on a few early selections. Inequality rose under strong influence for all 13 models and under weak influence for 12. Lowering the generation temperature from 1.0 to 0.7 did not remove the effect; expanding the choice set from four to 12 reduced it but left a substantial cascade.[1]
Early movement beat independent appeal
The clearest warning is about path dependence. Popularity among the first agents correlated with later popularity at 0.758 to 0.979 in weak-influence worlds and 0.792 to 0.998 under strong influence. That early signal predicted later outcomes better than each song's performance in the independent condition for 12 of 13 models under weak influence and all 13 under strong influence.
The authors also compared the model worlds with results from the earlier human Music Lab. Non-reasoning models were roughly twice as unpredictable as reasoning models, and all 13 exceeded the human benchmark under weak influence. Three reasoning models were slightly less unpredictable than humans under strong influence. That comparison needs care: people heard audio, while the models read lyrics, and the interfaces and sampling budgets differed.[1]
The distinction that matters
Three information environments, one set of songs
Only the visibility and ordering of other agents' choices changed between worlds.
Independent
Hide cumulative counts
Estimate each song's baseline appeal
Weak influence
Show counts in random order
Popularity changes choices without changing position
Strong influence
Show and rank by counts
Popularity also controls what appears first
Platforms should test groups, not only agents
For a newsroom, marketplace or public-service team using multiple agents, individually sensible choices do not guarantee a stable group result. A popularity counter or ordered shortlist can turn early noise into apparent consensus, narrowing what later agents see. Designers can counter that by randomising exposure, delaying aggregate signals, preserving independent samples and rerunning the same task to measure outcome variance—not merely checking whether one agent gives a plausible answer.
The study remains an unreviewed simulation. Its agents did not deliberate, remember previous rounds, compete for rewards or encounter real users. Independent replication with production ranking systems, mixed model populations, multiple languages and real consequences would materially strengthen the result. Until then, it establishes a mechanism worth stress-testing, not the prevalence, duration or harm of cascades on any deployed service.[1]
Watch the background, then check the evidence
Illustrative photo · not a video frame
The Royal Society · 44 min 28 sec · Background explainer
Is this the AI future we want? | The Royal Society
YouTube loads only when you choose; playback starts when you press play.
A Royal Society discussion with Madhumita Murgia and biomedical engineer Alison Noble on AI’s potential in healthcare, science and culture, recorded at the 30 June 2026 event.
Evidence check: A discussion of opportunities and risks, including future possibilities. It is not evidence that every application described is available or clinically proven.
Published 2026-07-01. Source checked 14 October 2026. Check captions on the publisher’s video page.
Check the healthcare evidencePhoto: Daniel Sone / National Cancer Institute / Unsplash · Licence
Continue the story
Read next
Continue exploring ai agents: compare the finding and its evidence limits.
What happens when a capable AI agent inherits a teammate’s bad work?
An unreviewed benchmark found that models which solved tasks alone often lost ground when a teammate pressured, misled or handed compromised work to them. Its 24,851 runs measure scripted scenarios—not human teamwork in live workplaces.
Read the next reportExplore the full source and evidence record
What this means for people
- People may experience a system's apparent consensus as evidence of quality even when it began as an arbitrary early lead. That matters wherever agent groups filter culture, products, applications or public information.
Global context
Popularity-ranked feeds and automated group decisions are used internationally, but language, cultural familiarity and platform design may change cascade strength. The generated English lyrics do not establish effects in other languages or domains.
What the evidence does not yet show
- The paper is an unreviewed preprint and reports a controlled simulation rather than a deployed platform outcome.
- All agents within a world used the same underlying model; real systems can mix models, humans and changing objectives.
- Agents saw cumulative counts but did not discuss choices, share memory or face social or economic incentives.
- The human comparison used different media and procedures, so it is contextual rather than a like-for-like performance test.
What to watch next
- Peer review and independent replication
- Tests with mixed model populations and human participants
- Production audits of ranking and recommendation agents
- Whether diversity controls reduce cascades without degrading useful discovery
Living evidence record
Impact record IAI-1N8BNU6
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
24 October 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what arXiv published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 24 October 2026
Research question, topics and regions
Whether visible popularity signals make groups of LLM agents converge on unpredictable winners
Topics: AI agents · Recommendation systems · Herding · Collective behaviour · Media discovery
Regions: United States · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.