Nature: AI is homogenising writing and possibly thought

When researchers tested 22 large language models and 102 humans on creative tasks, the models' answers were more similar to each other than the humans' answers were, even though individual model responses were rated slightly more original. The finding, from Duke University computer scientist Emily Wenger and a co-author in PNAS Nexus, sits at the centre of a Nature feature by Matthew Hutson rounding up a fast-growing body of work suggesting that generative AI narrows what people write, and possibly what they think.

An analysis of more than 400,000 papers in the Web of Science database found that after ChatGPT's late-2022 release, articles per author increased, and 'so did their similarity in content and linguistic style.' A separate linguistic study of local news, arXiv preprints and Reddit posts, led by University of Southern California PhD student Zhivar Sourati and published in Nature Human Behaviour, found that using LLMs for grammar correction 'erased many signifiers of personality, moral values and demographics.'

The effect appears to travel across cultures. In a 2025 CHI paper, Indian and American participants described favourite rituals, heroes and symbols with and without an AI autocomplete tool; the tool made their writing more similar and 'led Indians to sound more American in their wording,' with fewer details about cultural elements like Diwali.

It also appears to linger. In a study flagged by the piece, half of participants used ChatGPT for five days on creative tasks; when all were tested two months later without AI, those with previous access 'still gave answers that were more similar to each other.'

The named researchers do not soft-pedal the stake. 'The way that you talk affects the way that you reason,' Sourati tells Nature. Wenger frames it more bluntly: 'If we're all using AI to write our e-mails or whatever, what do we become as a species?' Sourati's worst case is that 'humans collectively become less adaptable. That's actually really scary for me.'

Hutson notes the phenomenon 'hasn't upended society and might never get to that point.' Two AI researchers we track had shared the feature within a day of publication.