Does AI Prefer Its Own Accent?

Does AI Prefer Its Own Accent?- A perspective on Self Preference Bias of models

A series on Agents and Biases

Part 1: A line I could not put down

In one of my session with Claude, it mentioned in passing that language models may favour text written by language models. It was almost an aside. I stopped reading.

Two questions came up and stayed. The first was practical: how would we check this? The second was stranger: how would a model even know? A piece of text does not arrive with a label. Some of it is written by a person. Some is written by a model. A growing share is written by a model and then tuned until it sounds like a person, or written by a person and then polished by a model. If the line between them is blurry to us, what exactly is the model responding to?

If AI agents are becoming the first reader of what sellers write, does it matter who, or what, wrote it?

Part 2: What the research made me infer

Models showed a preference for model-written text. In controlled tests, AI assistants were shown two descriptions of the same product or the same paper, one written by a person and one by a model, and asked to pick. They leaned towards the model-written version consistently.¹ The design borrowed from hiring discrimination studies: same item, different author.

People leaned the other way on products. In the same work, a small panel of human judges reading product descriptions blind chose the human-written one roughly twice as often as the model-written one.¹ So the model preference does not look like a simple case of "the AI text was better". On paper abstracts, people leaned towards the model versions too, which makes the product gap more interesting, not less.

In hiring simulations, the effect showed up on shortlists. When models screened résumés, they preferred résumés they had written themselves over human-written ones in 67% to 82% of comparisons, across commercial and open-source models. Candidates whose résumé came from the same model as the screener were 23% to 60% more likely to be shortlisted than equally qualified candidates with human-written résumés.²

Models can often tell their own writing apart, and that ability tracks the bias. Out of the box, some models distinguished their own outputs from other models' and from human text at better than chance. When researchers fine-tuned that recognition up, the self-preference rose with it, in a roughly linear way.³ In the hiring work, simple interventions aimed at the model's self-recognition cut the bias by more than half.²

But "knowing" may be the wrong word. Another line of work found that model judges gave higher ratings to text with lower perplexity, which roughly means text the model finds predictable, whether or not it wrote that text itself.⁴ That reframes my second question. The model may not need to recognise authorship at all. It may simply find some text easier to read, and easier to read may feel like better.

Which brings me back to the blurry line. Researchers working on AI-text detection argue that "LLM-generated text" has no consistent definition: human edits, model polishing, and models quietly shaping how people write all smear the boundary.⁵ If detectors struggle to draw that line, I doubt the reader-side preference respects it either.

The same familiarity also seems to help understanding. Researchers testing many prompts for the same task found that the lower a prompt's perplexity, the better the model performed.⁶ So text that is predictable to a model may be both easier for it to understand and more liked by it. From the outside, those two effects look the same.

Part 3: Familiarity, not identity

My working idea is that "does AI know it is AI-written?" may be a less useful question than "does this text sound like the reader?"

Accent premium (a thought experiment, not a finding): the extra weight an AI reader gives to text that sounds like something it could have written, separate from whether the content is better.

There is a human version of this. In studies of people, statements that are easier to process tend to be judged as more true.⁸ Behavioural economists call the wider family of effects processing fluency. I wonder whether a model has its own version of fluency, tuned to its own style instead of ours.

If that is roughly right, the useful variable is how close a text sits to the reader model's style, not who typed it. In my rough sketch:

Why this matters on the shelf

Amazon says independent sellers used its generative AI listing tools to create more than 12 million listings in 2025. In May 2026 it folded its Rufus assistant into Alexa for Shopping, which compares products from search results and can complete purchases, including on other retailers' sites through "Buy for Me". On at least one large marketplace, then, a growing share of the shelf is written by models and a growing share of browsing is done by one.

I read this as a quiet loop. Sellers use AI to write because it is cheap and fast. Agents read that text and may slightly favour it. And because the AI copy may also be cleaner, it is hard to separate quality from accent. I keep thinking of the shopkeeper from Issue 4 who writes his own descriptions in his own words. Is he paying a small tax he cannot see?

Should I write for the agent, then?

In my customer experience work, the first priority is making a product easy to understand. So my instinct was practical: if agents process predictable, AI-style text more easily, should product pages and interfaces be written that way to improve the agentic experience?

Part of the evidence says yes. In tests of generative search engines, adding quotations to content lifted its visibility in AI answers by about 41%, adding statistics by about 31%, and a fluency rewrite by about 28%, while keyword stuffing scored slightly below the untouched baseline.⁹ That reads to me like a reward for clarity, not for tricks.

But I think "agentic experience" bundles two things: being understood and being favoured. Making specs, prices and return rules clean and consistent probably helps an agent get your facts right. Rewriting your whole voice to sound like the model may only help you get picked. One design I keep coming back to is two layers on the same page: plain, structured facts for the agent, and the story in your own words for the person who reads after the agent has chosen.

And when almost everything is AI-written?

The studies compare human text with model text, but that contrast is fading as listings, help pages and code are increasingly drafted by models. My guess is the bias may then shift from human versus machine to typical versus unusual. When models were trained repeatedly on model-generated data, the rare parts of the original distribution disappeared first.¹⁰ That is a training finding, not a reading one, but it makes me wonder whether the unusual product and the local voice are the most fragile content in a machine-written world.

The questions I am sitting with

• If a model prefers text that is predictable to it, is that a bias or simply a reading style, and who gets to decide?
• When sellers "humanise" AI copy so it feels authentic to people, do they lose ground with agents, or gain it?
• If writing clearly for agents means sounding more like them, how much of a brand's own voice is worth keeping?

I write about AI, behavioural economics, and user research. Mostly questions, occasionally answers. If you are thinking about any of this, I would like to compare notes.

References

1. Laurito, W., Davis, B., Grietzer, P., Gavenčiak, T., Böhm, A., and Kulveit, J. (2025). AI-AI bias: Large language models favor communications generated by large language models. Proceedings of the National Academy of Sciences, 122. doi:10.1073/pnas.2415697122 (preprint: arXiv:2407.12856)
2. Xu, J., Li, G., and Jiang, J. Y. (2025). AI Self-preferencing in Algorithmic Hiring: Empirical Evidence and Insights. arXiv:2509.00462
3. Panickssery, A., Bowman, S. R., and Feng, S. (2024). LLM Evaluators Recognize and Favor Their Own Generations. NeurIPS 2024. arXiv:2404.13076
4. Wataoka, K., Takahashi, T., and Ri, R. (2024). Self-Preference Bias in LLM-as-a-Judge. arXiv:2410.21819
5. Geng, M., and Poibeau, T. (2026). On the Detectability of LLM-Generated Text: What Exactly Is LLM-Generated Text? ICML 2026 Workshop. icml.cc/virtual/2026/77477
6. Gonen, H., Iyer, S., Blevins, T., Smith, N. A., and Zettlemoyer, L. (2023). Demystifying Prompts in Language Models via Perplexity Estimation. Findings of EMNLP 2023. arXiv:2212.04037
7. Chen, W.-L., Wei, Z., Zhu, X., Feng, S., and Meng, Y. (2025). Do LLM Evaluators Prefer Themselves for a Reason? arXiv:2504.03846
8. Reber, R., and Schwarz, N. (1999). Effects of perceptual fluency on judgments of truth. Consciousness and Cognition, 8(3), 338 to 342.
9. Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., and Deshpande, A. (2024). GEO: Generative Engine Optimization. KDD 2024. arXiv:2311.09735
10. Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R., and Gal, Y. (2024). AI models collapse when trained on recursively generated data. Nature, 631. doi:10.1038/s41586-024-07566-y

Sources

• PYMNTS (30 April 2026), Amazon Sellers Gain Sales and Cut Costs With AI Tools: pymnts.com
• EcomCrew (May 2026), Amazon Retires Rufus and Launches Alexa for Shopping: ecomcrew.com
• Semafor (24 April 2026), Google CEO says 75% of company's new code is AI-generated: semafor.com

LinkedIn version

I've started to wonder whether AI shoppers have an accent preference.

Not for who wrote a product page, but for whether it sounds like them.

A throwaway line from Claude stopped me this week: models may favour text written by models.

In controlled tests, AI assistants choosing between two descriptions of the same product leaned towards the model-written one. A small human panel leaned the other way.

In hiring simulations, models preferred résumés they had written themselves in 67% to 82% of comparisons.

The part that stuck with me: the models may not need to "know" who wrote anything. One line of research found model judges rated text they found more predictable higher, whoever wrote it.

That sounds a lot like processing fluency in people. Easy to read feels true.

It also raised a question from my own CX work: if predictable text helps agents understand us, should we write for them? For clean facts, I think yes. For rewriting our whole voice to sound like them, I'm less sure.

And in code, where Google's CEO says about 75% of new code is AI-generated, at least there are tests to referee. Product text has none.

Now the shelf. Amazon says sellers used its AI listing tools to create more than 12 million listings in 2025, and its new Alexa for Shopping can compare and buy for you.

Machines writing the shelf. Machines reading it.

In my rough sketch, the seller who still writes by hand may be paying a small tax they cannot see.

I don't think the useful question is "does AI prefer AI?" It may be: where does the agent's preference stop matching what the person wanted?

If you run research on product pages: have you tested the same pages on an agent and on people? What did the gap look like?

#AgentsAndBiases #AgenticCommerce #BehavioralEconomics

First comment: The full post, with references and a simple test you can run in an afternoon: [BLOG LINK]

Slides (4 x 1080 square, Cover + Parts 1 to 3): Issue 6 slides

Input (internal)