How do I keep my own voice when using AI?

Voice erodes without anyone deciding to erode it, and the evidence says the loss runs deeper than style. In a study of 1,506 people, writing alongside a model configured to hold an opinion changed not only what participants wrote but what they thought afterwards, measured on a survey taken after the writing was done. The threat to your voice is that the position underneath it has moved, rather than that the prose sounds generic.

The finding about opinions, not style#

Jakesch and colleagues gave 1,506 participants a writing assistant while they composed a post on whether social media is good for society. The assistant was configured to argue one way or the other. Five hundred independent judges rated the opinions expressed, and participants then completed an attitude survey.

The model shifted both the opinions expressed in the writing and the participants' own opinions afterwards. The effect held among people who had plenty of time to write independently, which rules out the easy explanation that they accepted suggestions to save effort. The authors call it latent persuasion.

One topic, one configuration, self-reported attitudes, and no test of whether the shift persisted. It is one study and it should not be over-read. It is also the only study anyone has run that asks the question directly, and the direction it points is not reassuring.

Individually better, collectively narrower#

The best-known result here is Doshi and Hauser's, published in Science Advances. Three hundred writers produced short stories, with no AI, one AI idea, or up to five. Six hundred judges rated them.

Both halves matter and only one usually gets quoted. Individual creativity rose, and rose most for the writers who had scored lowest on an independent creativity measure, by up to 26.6 per cent on how well written the stories were. And the AI-assisted stories were measurably more similar to one another, around 10.7 per cent more similar in the single-idea condition.

Everyone got better. Everyone got better in the same direction. That is a social dilemma rather than an individual failure, so trying harder on your own cannot solve it.

A 2026 analysis of 6,875 essays found the same trade in more detail, and complicated it usefully. Structural features converged sharply, with cohesion architecture losing 70 to 78 per cent of its variance. But perspective plurality actually diversified. So the flattening is not uniform. The shape of the writing converges while the range of positions taken need not, which suggests what to protect and what to stop worrying about.

Ownership tracks how much of you is in it#

Joshi and Vogel measured something more personal: whether the work still feels like yours. Participants wrote short stories under conditions ranging from a three-word prompt to writing unaided.

Psychological ownership rose steadily with prompt length, from a mean of 1.80 with a three-word prompt to 6.29 writing alone. The gain plateaued once the prompt reached roughly the length of the finished piece, and no AI-assisted condition ever reached the ownership of writing unaided.

There is a practical reading. The more of the thinking you put into the prompt, the more the result is yours, up to a point past which you may as well have written it. And if authorship matters for a particular piece, no amount of prompting recovers what writing it yourself gives you.

The question nobody has answered#

Can people tell that their own output has become more generic?

This research could find no study that tests it. There is work on whether readers can spot machine text, and work on how detection anxiety changes writing behaviour, but nothing measuring whether a writer notices the flattening in their own work.

That gap matters more than it looks. Every self-management strategy on this page assumes you would notice. The Doshi and Hauser design had to use embeddings and independent judges to see the convergence, because it is a property of the collection rather than of any single piece. From inside your own document, a more conventional version of your idea reads as a cleaner version of your idea.

A wider synthesis published in Trends in Cognitive Sciences argues that models reflect and reinforce dominant styles and that reliance on a small number of systems amplifies convergence. It is a perspective piece with no new measurement of its own, and is cited here as a statement of concern across several fields rather than as evidence.

What to do about it#

• Form the view before the tool sees it. The opinion-shift finding is the serious one, and it operates through exposure. A position you wrote down first is one you can notice moving.
• Put the thinking in the prompt, not the polish. Ownership tracks how much of you went in. A three-word prompt produces something that is not yours in any sense you would defend.
• Decide which pieces are authorship pieces. For some writing, being the author is the whole point. For those, the evidence says there is no prompting strategy that substitutes.
• Protect the position, worry less about the prose. The essay analysis suggests structure converges hardest while perspective need not. Your sentence rhythm is the smaller loss.
• Assume you will not notice. Nobody has shown that people detect their own homogenisation, and the effect is only visible across many pieces. Keep unaided samples so there is something to compare against.

Related SuperSkills research#

On the collective version of this, does AI make everyone think alike. On forming the view first, Human at the Start. On agreement as the other route to the same place, how to get AI to challenge you. On what is lost more broadly, using AI without dependency and what stays human. On the expression of noticing another person, outsourced recognition.

Key research and primary sources

• Jakesch, M. et al. (2023). Co-Writing with Opinionated Language Models Affects Users' Views. CHI 2023.
• Doshi, A. R. and Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28).
• Joshi, N. and Vogel, D. (2025). Writing with AI Lowers Psychological Ownership, but Longer Prompts Can Help. ACM CUI 2025.
• Sourati, Z., Ziabari, A. S. and Dehghani, M. (2026). The Homogenizing Effect of Large Language Models on Human Expression and Thought. Trends in Cognitive Sciences.

About this research#

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. The opinion-shift result rests on a single study with one topic and one configuration, which the page states. Whether writers can detect their own homogenisation has not been studied at all, and the page says so rather than assuming the answer.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.