AI as a source of ideas: helpful individually, more similar collectively

What the story experiment actually showed

Anil Doshi and Oliver Hauser conducted an online experiment examining how ideas from a language model influence the writing of short stories. Some participants received such suggestions; others did not. In the setting studied, stories produced with AI ideas were rated as more creative, better written and more enjoyable. People with lower measured baseline creativity benefited particularly clearly. At the same time, the AI-assisted stories were more similar to one another. The result thus describes two levels that are easily confused: the success of an individual text and the diversity of an entire collection. It is not proof that AI makes all creative work more uniform or generally improves it. The study examined a particular task, specific forms of assistance and the subsequent ratings within this experimental setup, not every form of artistic collaboration.

A good suggestion can also become an anchor

Someone who receives a fully developed idea early on may follow it more closely than initially intended. The assistance then reduces the emptiness of the blank page, but also limits the space for exploration. This is particularly relevant when many people use similar systems with similar requests. The study highlights this trade-off without thereby prescribing every creative routine. For individual projects, a familiar pattern may well be desirable. In a collection of stories, campaigns or designs, however, distinctiveness is often precisely what matters. The usefulness of an assistant should therefore not be measured solely by whether an individual suggestion is immediately appealing. It is equally revealing to consider which ideas of one's own are consequently not pursued, and whether several results end up repeating the same motifs, conflicts or turns of phrase.

A way of working that preserves independence

One possible workflow begins with an outline of one's own before AI is brought in. It is then given a limited role: finding counterarguments, identifying blind spots or suggesting alternative structures. Rather than adopting a finished draft, several distinctly different directions are compared. Teams can initially work independently of one another and only use AI assistance in a later round of revision. This does not guarantee that uniformity will be prevented, but it creates a more deliberate separation between one's own initial idea and the machine's suggestion. It is also helpful to keep rejected alternatives and explain decisions. This makes it possible to trace why a text took the form it did. The central question is not whether AI is creative, but whether the specific way it is used supports a project's intended independence and diversity.

Keep exploring

• Interpreting AlphaFold correctly: structure predictions need uncertainty information
• AI image descriptions: Useful when the purpose is clear
• Data quality: why provenance and selection matter more than sheer volume