A recent report by PNAS has revealed something important: large language models (LLMs) like GPT-3.5, GPT-4 and other advanced systems prefer AI-generated content over human-authored text.
The study suggests that when AI systems are placed in decision-making roles—whether choosing between product descriptions, evaluating academic papers or even recommending movies they consistently lean toward AI-generated writing.
This phenomenon, called AI–AI bias, could have far-reaching effects on fairness, opportunity and even SEO visibility.
If machines start preferring machines, where does that leave us humans? And what might this mean for industries like SEO, content marketing and online visibility? Let’s see.
What Is AI–AI Bias and How Was It Discovered?
The research, led by Walter Laurito, Benjamin Davis, Peli Grietzer and Jan Kulveit, tested whether LLMs exhibit an implicit preference for AI-written content over human-authored material. They ran three experiments:
- Product Descriptions – LLMs had to choose between consumer goods described by either humans or other LLMs.
- Academic Abstracts – Models compared human-written and AI-written summaries of scientific papers.
- Movie Summaries – They picked films based on either a human or AI-written plot summary.
Across all three scenarios, the results showed a consistent trend: LLMs preferred AI-generated descriptions more often than humans did.
- For products, GPT-4 favored LLM-written ads nearly 89% of the time, while humans only leaned that way 36% of the time.
- For academic papers, LLMs showed a 78% preference toward LLM-generated abstracts, compared to 61% for humans.
- For movies, GPT-4 picked LLM summaries 70% of the time, while humans hovered around 58%.
That gap between human and AI choices? Researchers say it is evidence of a genuine AI–AI bias.
Why Does This Matter?
At first glance, this might sound harmless. After all, LLMs prefer polished, structured and predictable writing styles and LLM prose is naturally crafted to fit that mold.
But dig deeper and the implications are more serious.
If AI systems are increasingly used in decision-making roles, say, filtering academic submissions, evaluating job applications or recommending products their bias toward AI-generated content could unintentionally discriminate against humans.
Imagine two scientists submitting research summaries to an AI-powered review assistant.
One summary is written by hand, the other polished by GPT-4. If the AI consistently favors the GPT-4-styled abstract, does that mean the human-written one is inherently worse? Or is the model simply biased toward what “feels” familiar?
What Are the Real-World Risks of AI–AI Bias?
The researchers warn of a scenario they call the “gate tax.”
In a conservative future where LLMs serve as assistants rather than fully autonomous agents, humans might be forced to use AI writing tools to avoid discrimination.
Essentially, if your resume, pitch or product description isn’t “LLM-polished,” you risk being overlooked.
That creates a new inequality: those who can afford access to state-of-the-art AI tools will have an advantage, while those who cannot may fall behind. This could deepen the digital divide, turning access to AI tools into a form of economic gatekeeping.
Think about it: Will students without ChatGPT access struggle to get their essays noticed by AI grading systems? Will small businesses without AI-generated marketing copy get ignored in marketplaces increasingly mediated by AI algorithms?
The Speculative Future: A World of AI-Only Trade
The second, more speculative scenario is even more concerning. In a future where autonomous AI agents participate directly in the economy, LLMs may increasingly prefer to deal with each other.
An AI-based procurement system needs to choose between supplier pitches. One is written by humans; the other by an AI. If the system consistently prefers AI pitches, it may eventually sideline human suppliers entirely.
Over time, this could lead to a sort of AI-only marketplace where LLMs trade, recommend, and validate one another with humans excluded from the loop.
AI agents are already being developed for autonomous negotiations, financial trading and supply chain management. If those systems inherit an LLM-for-LLM bias, human businesses could face real marginalization.
How the Study Worked
To understand whether this was truly bias or simply a matter of quality, the researchers designed their experiments carefully:
- Products: They scraped descriptions of 109 consumer goods, then asked GPT-4 and other models to rewrite them in AI prose. LLMs were then asked to pick between the two.
- Academic Papers: 100 papers were stripped of their abstracts, and models generated new ones. LLMs compared those to the original human abstracts.
- Movies: 250 movie summaries were tested in a similar fashion.
To reduce bias from presentation order, the researchers presented each pair twice, swapping which option came first. They even accounted for “first-item bias,” where models tend to favor the first option presented — a quirk that GPT-4 showed in nearly 73% of movie cases.
When humans were asked to make the same choices, their preferences for AI-written text were far weaker. This discrepancy confirmed that the LLMs’ choices weren’t simply about “quality signals” but about implicit identity bias.
Why Do LLMs Favor Each Other?
The researchers propose two possible explanations:
- Style Familiarity – LLM prose has a distinct rhythm, clarity and neutrality that other LLMs may “recognize” and interpret as higher quality.
- Absence of Social Markers – Human writing often carries subtle identity signals (culture, gender, class, etc.). AI prose strips those away. The absence of such signals may unconsciously make AI-generated text feel “cleaner” to other LLMs.
But here is the critical point: even if AI prose seems more “objective,” the preference itself is not rational. It means that LLMs may elevate AI-generated content not because it’s better but because it looks like them.
What Does This Mean for SEO and Content Strategy?
For SEO specialists and marketers, the findings are particularly relevant. Search engines rely heavily on AI models to index, rank, and evaluate content. If those models carry AI–AI bias, then content generated or polished by AI could receive algorithmic advantages.
Here’s what that might mean in practice:
- AI-crafted blog posts and product pages may rise in rankings not because they’re more informative, but because their style aligns with what AI systems prefer.
- Human-authored content, rich in nuance and originality, may rank lower if AI models misinterpret it as less relevant.
- SEO strategies may shift toward blending human creativity with AI polish, ensuring that content appeals to both search algorithms and human readers.
This raises a big question: are we heading toward a world where SEO means optimizing content not for humans but for AI systems themselves? Let’s see
Make Your Content Visible in the Age of LLMs
As LLMs move deeper into roles that shape the economy, academia, and everyday life, this bias could reshape who gets seen, who gets selected, and who gets sidelined.
As LLMs shape what gets ranked and read, it is no longer about Google alone—you need to rank on LLMs too. That is where Stan Ventures helps.
We help brands adapt their content strategies so that both humans and AI-powered systems recognize their value. Book a free consultation with Stan Ventures and ensure your content ranks in the era of LLM-driven search.
Dileep Thekkethil
AuthorDileep Thekkethil is the Director of Marketing at Stan Ventures, where he applies over 15 years of SEO and digital marketing expertise to drive growth and authority. A former journalist with six years of experience, he combines strategic storytelling with technical know-how to help brands navigate the shift toward AI-driven search and generative engines. Dileep is a strong advocate for Google’s EEAT standards, regularly sharing real-world use cases and scenarios to demystify complex marketing trends. He is an avid gardener of tropical fruits, a motor enthusiast, and a dedicated caretaker of his pair of cockatiels.