Anyone who reads a lot of email has noticed the shift. Messages arrive with the same balanced sentence rhythms, the same confident summary paragraph, the same three-part lists. Two colleagues who never wrote alike now do.
The research term for this is homogenisation: when different people use the same AI assistant, their outputs converge. It is one of the better-replicated findings in the writing-with-AI literature, and it sits next to a finding that looks like its opposite. The same experiments show that AI assistance makes each individual piece rate better. Both are true at once, which is why the honest framing is a trade rather than a verdict: your output improves, the distinctiveness of it, and possibly of you, goes down.
The vocabulary you are seeing
Two terms are circulating, and they mean different things.
“Slop” was Merriam-Webster’s 2025 word of the year, defined as digital content of low quality that is produced usually in quantity by means of artificial intelligence. It describes the flood: junk videos, spam books, filler posts. It is a judgement about quality and volume.
“Homogenisation” is the research term and it is narrower. It describes a measured reduction in the diversity of what a group of people produce when they use the same tool, regardless of whether any single piece is good. AI-assisted writing can be well above average and homogenised at the same time. That combination is what the experiments keep finding.
The core experiment: better stories, more alike
The cleanest test is Doshi and Hauser’s 2024 paper in Science Advances. In an online experiment, writers produced short stories either alone or with access to story ideas from a large language model. Independent evaluators rated the AI-assisted stories as more creative, better written and more enjoyable, and the boost was largest for writers who scored as less creative to begin with.
Then the authors measured how similar the stories were to one another. Stories written with AI ideas were more similar to each other than stories written by humans alone. The authors describe the result as a social dilemma: each writer is individually better off using the tool, and the collection of stories they produce together is narrower.
That structure, individual gain and collective loss, is the pattern to carry into every other claim in this article. It explains why people who use these tools sincerely report that their writing improved, and why readers of many such people sincerely report that everything sounds the same. Both observations are accurate.
Ideas narrow too, and ownership drops
A 2024 study by Anderson, Shah and Kreminski at the Creativity and Cognition conference asked the same question about ideas rather than prose. Thirty-six participants generated ideas either with ChatGPT or with a non-AI creativity support tool. The ChatGPT group produced more ideas, and more detailed ones. Across participants, though, those ideas were less semantically distinct from one another: different people using the same model converged on similar suggestions. The ChatGPT users also reported feeling less responsible for the ideas they had generated.
That last measurement, ownership, is the bridge between output and writer. A tool that raises your output while lowering your sense that it is yours changes what you have practised.
The vocabulary evidence shows the convergence at scale. Kobak and colleagues, in Science Advances in 2025, analysed more than 15 million biomedical abstracts indexed in PubMed from 2010 to 2024 and tracked abrupt rises in the frequency of particular style words after language models became available. Their estimate is that at least 13.5% of 2024 abstracts were processed with a language model, a lower bound that reached 40% in some subgroups. The words that spiked are the ones you have started noticing in your inbox, and the effect on scientific vocabulary exceeded that of the COVID pandemic.
What happens to the writer’s own skill
This is the question people search for and the one with the least evidence. What exists is small.
The most discussed study is Kosmyna and colleagues’ 2025 MIT Media Lab preprint, often called the “cognitive debt” paper. Fifty-four participants wrote essays over three sessions using either a language model, a search engine or no tool, with EEG recording, and 18 returned for a fourth session with tools swapped. The language model group showed the weakest brain connectivity, reported low ownership of their essays in interviews, and struggled to quote from essays they had written minutes earlier. The paper also found within-group homogeneity in named entities, phrasing and topics.
Read it with its critics. A commentary posted in 2026 by Stanković and colleagues raises the limited sample size, the reproducibility of the analyses and methodological issues in the EEG work, and argues some results should be interpreted more conservatively. The preprint is not peer reviewed. What it offers is a plausible mechanism with a small sample and a loud reception, which is not the same as a demonstrated decline in writing skill.
Beyond that, there is no longitudinal study of whether people who write with AI for a year write worse without it. There are interview accounts and the ownership measures above. Anyone telling you the skill loss is established is ahead of the data, and so is anyone telling you it is not happening.
The balance, stated plainly
On the evidence, three things are supported. Individual output rated by others improves, especially for weaker writers. Outputs across people become more similar, in prose, in ideas and in vocabulary. Writers using the tool feel less ownership of what they produce.
Two things are not supported yet. That the writer’s underlying skill measurably degrades; the studies are small and short. And that the homogenisation is permanent or unavoidable; the experiments compare a single session with and without the tool, not a person who has learned to push against the default.
Common questions
What is AI homogenization?
It is the measured finding that when different people use the same AI assistant, what they produce becomes more similar to one another. In one experiment, short stories written with AI ideas were rated as better and were more alike than stories written unaided; in another, 36 people using ChatGPT produced ideas that were less distinct from each other than people using a non-AI tool.
Does AI make your writing worse?
Not by the ratings. Evaluators judged AI-assisted stories more creative, better written and more enjoyable, with the largest gain for less creative writers. What declines is distinctiveness across writers and, in the studies that measured it, the writer’s own sense of ownership. Whether your unaided skill degrades over time has not been measured beyond one small preprint.
What is AI slop?
Merriam-Webster chose “slop” as its 2025 word of the year, defining it as digital content of low quality that is produced usually in quantity by means of artificial intelligence. It refers to volume and quality. Homogenisation is a separate research finding about similarity, and can apply to content that is not low quality at all.
How common is AI-assisted writing?
In biomedical research, an analysis of more than 15 million PubMed abstracts estimated that at least 13.5% of 2024 abstracts were processed with a language model, rising to 40% in some fields and countries. That is a lower bound based on vocabulary shifts, and it excludes lighter editing that leaves no trace.
The bottom line
If you use a model to draft, expect your work to be rated a little better and to sound a little more like everyone else who does the same. Both effects are measured. The risk to your own voice is real in the sense that you will feel less ownership of what you send, and unproven in the sense that nobody has shown your unaided skill declines. The practical response follows from the mechanism: keep drafting something yourself before you ask for help, treat the model’s first suggestion as the average of everyone’s suggestion, and edit against it rather than towards it. The collective narrowing is an experiment result. Whether you personally narrow is still up to you.
Did this answer your question?
Thanks. That helps us decide what to fix.
Leave a comment