We Trained AI on Our Language. Now It’s Shaping Ours.

We use language to communicate ideas, but we also use it to discover them.

Anyone who spends a lot of time thinking, writing or doing creative work has probably experienced this. You can have an idea that feels clear in your head, then try to put it into words and struggle. You know roughly what you mean, you can almost feel where the idea is going, but you can’t seem to articulate it.

Sometimes struggling with the words is how you realize that the idea itself isn’t completely formed yet. You write, go back, change something, try another way of saying it, and somewhere in that process you understand more clearly what you were actually trying to say. Rewriting changes more than the actual wording. It can make you understand the idea differently by the time you finish the paragraph.

You can use generative AI at almost any point in this process. You can use it to research a subject, go through a large amount of material, summarize it, test an argument, connect ideas, work through a draft or simply help you find the words. A model can point you to a connection you missed or put into words something you already had a mental understanding of but couldn’t yet articulate.

It can also give form to an idea before you have fully defined or refined it for yourself. And when that version comes back more fluent, coherent and ready to use, it can influence the direction you take before you have completely decided for yourself that you agree with that direction.

Once you start going back and forth with the model, asking for examples, alternatives, connections or different ways of structuring something, it becomes harder to tell which parts were already clear to you and which developed through the exchange. You may bring the question, the research and the judgment, then receive a formulation you probably wouldn’t have reached by yourself, or at least not as quickly. You can also accept a different version of your idea simply because it reads well and is easier to use than to keep searching for your own wording.

Millions of people are now using AI in some version of this way every day, from work and research to coding, planning and solving all sorts of problems.

Human language went into these models first. They were trained on our writing and ideas, on our ways of explaining things. Now generated language is coming back into the same environment people read, learn, work and write in. And this exchange, which at the moment goes unnoticed by most people, affects far more than what we call style.

What happens to individual expression when the same kinds of systems help large numbers of people formulate it? What happens to discernment and critical thinking when fluent language becomes cheap and immediate? How much of the work of reaching a certain way of thinking can we hand over before this convenience starts changing the way we actually think?
Of course, AI can help you make something better. But when a lot of people are getting help from similar models, the work they produce can also start to look too much alike.

In a 2024 experiment, researchers Anil Doshi of UCL School of Management and Oliver Hauser, then at the University of Exeter, asked 293 people to write short stories, with some participants given ideas generated by GPT-4. The AI-assisted stories were rated as more creative, better written and more enjoyable, especially when the writers themselves had scored lower on creativity beforehand. But when the researchers compared the stories, they found that the AI-assisted ones were a lot more alike. So, for the person using AI, it might just seem like a good result because the story came out better. Across the group, though, there was less variety, and that only becomes visible when the stories are looked at together.

There was also a much larger study, published in Nature Human Behaviour in August 2026, that looked at more than 880,000 texts across seven datasets and several kinds of writing.

The researchers found that as large language models became more common as writing assistants, linguistic diversity went down.

They also ran experiments where models were asked to polish or rewrite people’s text. Most of what people were saying stayed the same, but the differences in how they wrote became smaller. Depending on the dataset and the model, the variance dropped by 21% to 50%. The rewriting even altered some of the linguistic cues researchers can use to infer things like a writer’s age, ideology or moral values.

This doesn’t mean everyone suddenly starts sounding the same. What seems to happen is that some of the differences in how people write begin to fade, even when they are still saying essentially the same thing.

Language can’t be reduced to the information being communicated. Two people can say almost the same thing and still sound completely different depending on the tone, word choice and rhythm. Even the sense of the person behind the text can change while the basic content stays pretty much the same.

For a private task, none of this has to be a problem. If AI helps you organize notes or solve something that stays between you and the tool, it probably doesn’t matter if the phrasing resembles someone else’s. But it does matter when that language and those materials start circulating. Because then other people read it, absorb it, learn from it and sometimes, consciously or not, they reproduce it and spread it.

The term AI slop is often used for the flood of low-effort, machine-generated content now appearing online. A lot of it’s just… bad. But there’s also the sheer amount of it. When huge volumes of generated content enter public spaces, they become part of what other people read, watch, reference and learn from.

We wrote the material these models learned from, and now their language is, in turn, becoming part of our language. That exchange is already happening in creative work, professional writing, education and public communication, even if most of us rarely stop to think about it.

Using AI well also means noticing what we’re handing over. Sometimes delegating is the wisest choice we could make. There’s no virtue in spending an hour on a task a model can do well in two minutes. But there are other cases when the difficult part is the part we need to tackle ourselves. Trying to find the words can help us realize that we don’t fully understand the idea yet. It can push us to make a certain distinction we were avoiding or show us a lack in our argument.

AI literacy includes knowing how to prompt, but it also includes knowing if the final direction still reflects your own judgment. AI can save us hours of works. But if you’re still trying to understand what you think, spending the time may be exactly what helps you get there. And maybe part of AI literacy is simply getting better at knowing when to save the hours and when to spend them.

Series note

This article is part of the SourceLess AI Series, an ongoing look at AI from technical, social and human perspectives, and at how these systems are changing the way people work, create, communicate and make decisions.

Sources

• Anil R. Doshi & Oliver P. Hauser (2024). ‘Generative AI enhances individual creativity but reduces the collective diversity of novel content.’ Science Advances 10(28), eadn5290. DOI: 10.1126/sciadv.adn5290.
• Zhivar Sourati et al. (2026). ‘The shrinking landscape of linguistic diversity in the age of large language models.’ Nature Human Behaviour. Published 24 August 2026. DOI: 10.1038/s41562-026-02550-0.