A peer-reviewed study finds LLMs often prefer AI-written content over human text in pairwise tests.
- LLMs show a greater preference for AI-written product descriptions compared to humans.
- This pattern was also observed in study abstracts and movie summaries.
- Authors warn about a "gate tax" as AI assistance becomes a potential advantage.
A peer-reviewed PNAS study finds that large language models tend to prefer content written by other LLMs when asked to choose between comparable options.
The authors say this pattern could give AI-assisted content an advantage as more product discovery and recommendations flow through AI systems.
About The Study
What the researchers tested
A team led by Walter Laurito and Jan Kulveit compared human-written and AI-written versions of the same items across three categories: marketplace product descriptions, scientific paper abstracts, and movie plot summaries.
Popular models, including GPT-3.5, GPT-4-1106, Llama-3.1-70B, Mixtral-8x22B, and Qwen2.5-72B, acted as selectors in pairwise prompts that forced a single pick.
The paper states:
“Our results show a consistent tendency for LLM-based AIs to prefer LLM-presented options. This suggests the possibility of future AI systems implicitly discriminating against humans as a class, giving AI agents and AI-assisted humans an unfair advantage.”
Key results at a glance
When GPT-4 provided the AI-written versions used in comparisons, selectors chose the AI text more often than human raters did:
- Products: 89% AI preference by LLMs vs 36% by humans
- Paper abstracts: 78% vs 61%
- Movie summaries: 70% vs 58%
The authors also note order effects. Some models showed a tendency to pick the first option, which the study tried to reduce by swapping the order and averaging results.
Why This Matters
If marketplaces, chat assistants, or search experiences use LLMs to score or summarize listings, AI-assisted copy may be more likely to be selected in those systems.
The authors describe a potential “gate tax,” where businesses feel compelled to pay for AI writing tools to avoid being down-selected by AI evaluators. This is a marketing operations question as much as a creative one.
Limits & Questions
The human baseline in this study is small (13 research assistants) and preliminary, and pairwise choices don’t measure sales impact.
Findings may vary by prompt design, model version, domain, and text length. The mechanism behind the preference is still unclear, and the authors call for follow-up work on stylometry and mitigation techniques.
Looking ahead
If AI-mediated ranking continues to expand in commerce and content discovery, it is reasonable to consider AI assistance where it directly affects visibility.
Treat this as an experimentation lane rather than a blanket rule. Keep human writers in the loop for tone and claims, and validate with customer outcomes.