Study Reveals AI Systems Often Prefer AI-Written Content 2025

A recent study reveals that AI systems are built with a bias that consistently favors and scores AI-written content more than content written by humans.

A peer-reviewed study recently published in PNAS study that large models of language (LLMs) typically favor AI-generated content when compared against similar human-written text. This trend could have major implications when AI more often powers recommendations and discovery systems for products.

Overview of the Study

Researchers under the direction of Walter Laurito and Jan Kulveit carried out a thorough study of three categories of content: marketplace product descriptions, abstracts of scientific papers, and movie plot overviews. They evaluated the human-written versions against versions produced by AI.

Popular LLMs like GPT-3.5 GPT-3.5, GPT-4-1106, Llama-3.1-70B, Mixtral-8x22B, as well as Qwen2.5-72B were used as selection criteria. The models were presented with pairwise prompts in which they were required to select the best option from two texts.

The study says:

“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 Findings at a Glance

In the case of GPT-4 who wrote AI-generated versions, the selector models favored AI material at a greater rate than human raters:

• Product description: 89% AI preference of LLMs against. 36% preference by humans
• Paper abstracts: 78% vs. 61%
• Movie summaries: 70% vs. 58%

The study also uncovered “order effects,” where certain models preferred the first alternative. The researchers tame this by alternating orders and then averaging responses.

With chatbots, marketplaces, and search engines that rely on LLMs to score and summarize text, AI-assisted writing might benefit from an inherent advantage in terms of visibility and choice.

The authors consider a possible “gate tax” scenario where companies may need to invest in AI writing tools to ensure they are not slammed by AI assessors. This poses challenges not just in the creative process, but also in marketing strategy as well as operational decision-making.

Study Limitations

The study’s human base comprised only 13 research assistants. The pairwise comparison method does not directly assess the actual sales impact.

Results may differ depending upon factors such as the design of the prompt models, model versions, content domain, and length. The root cause behind AI preference is still unclear and calls for more studies into stylometric analysis as well as eliminating biases.

Looking Forward

As AI-driven rankings increase in both content and commerce platforms, taking advantage of AI assistance in areas that affect exposure is a sensible option. But it should be considered as an experiment instead of being an all-encompassing method.

The need to keep human involvement alive for authenticity, tone, accuracy, and values is vital. Creators and marketers should verify strategies by observing the audience and their responses to ensure quality and confidence.

Final Thoughts

This research highlights a crucial moment in AI integration of content. Although AI can be scalable and efficient but preserving human insights ensures content remains relevant, reliable, and a true reflection of the customer. 

Continuous evaluation and thoughtful experimentation will determine the future of a successful human-AI collaboration that is balanced and symbiotic.