The stories grew alike. Each one got better.

Last case, and the one I’d frame on the wall. Doshi (UCL School of Management) and Hauser (University of Exeter) published in Science Advances in 2024: “Generative AI enhances individual creativity but reduces the collective diversity of novel content”. Peer reviewed, and a preregistered experiment, which means they said what they’d test before they tested it. Rare. Lovely.

The setup: 293 writers wrote eight-sentence stories. Some writers got no AI help at all, while others got one GPT-4 story idea or up to five. Then 600 evaluators made 3,519 evaluations, and similarity was measured with text embeddings.

The good news first. With up to five AI ideas, novelty rose 8.1% and usefulness 9.0%. Less creative writers gained the most: 10 to 11% on creativity and 22 to 26% on how enjoyable and well written the story was. If you’ve ever coached a junior who freezes at a blank page, you can see the appeal.

But here’s the catch. The stories became more similar to each other. The increase in similarity from one or five AI ideas was 10.7% and 8.9% of the total range. Their summary says it best: with generative AI, writers are individually better off, but collectively a narrower scope of novel content is produced.

That’s the whole theory on trial, in miniature. Everyone’s story improves and everyone’s story starts to rhyme.

Caveats, honestly given. It’s a lab experiment with short fiction, early GPT-4, and AI ideas rather than AI-written text. And the similarity effect is modest; nobody should read this as proof the world is about to sound identical.

Still. A tool that makes each of us a bit better and all of us a bit closer is exactly the sort of thing you’d miss from inside your own inbox.

Verdict: split.