Five good AI ideas can still be the same idea
Higher-rated drafts can become more alike. Keep independent observations before asking AI to generate, refine, or group creative options.

Five individually good ideas can still leave a team with one idea repeated five times. That is the risk I would look for when everyone starts a creative task by asking an assistant for suggestions.
The output may be useful. It may even receive higher ratings than an unaided draft. The question is whether the collection still contains different starting points, assumptions, and ways of seeing the problem.
That is a more defensible concern than saying AI has killed creativity. It also suggests a testable change to the workflow: keep some independent ideas before introducing a common source of suggestions.
Individual quality and collective variety are different outcomes
In the study later published in 2024, Doshi and Hauser gave some writers access to AI-generated story ideas. The resulting stories received better evaluations, especially for writers with lower initial creativity measures. They were also more similar to one another than the unaided stories.
These findings can coexist. A useful suggestion may improve one person's story while drawing several people towards a similar direction. The study concerns a particular creative-writing experiment, not every design task or the commercial value of every AI-assisted product.
A 2026 experiment with 302 university students also found that ChatGPT access improved performance on creative picture-idea tasks. Benefits varied with participants' baseline creative potential. The researchers did not find evidence that the human-AI teams outperformed the model alone through complementarity in that setting.
That result does not tell a designer to stop using the tool. It does make a vague claim that collaboration automatically combines the best of both sides difficult to sustain. The task, comparison, and evaluation criterion decide what improvement means.
Keep the first observations independent
I would start a team exercise by asking each participant to record a few observations from the actual problem before seeing generated suggestions or one another's answers. These might be a user difficulty, an unusual constraint, or an example that contradicts the obvious approach.
Save those notes. Then introduce the assistant for a defined job: propose variations, find missing assumptions, or explain where an idea might fail. Keeping the original notes makes it possible to see which directions survived and which disappeared after the model entered the process.
A July 2026 preprint tested creative metaphors with native and non-native English writers. In that experiment, AI-generated starting ideas reduced collective diversity, while AI refinement of participants' own ideas preserved it. This is early evidence from a bounded task, not a guarantee for all creative work.
It gives the sequence a reason to test. Generating the starting point and editing an existing one are different interventions. A team should not treat them as interchangeable because both involve the same chat interface.
Compare the set before choosing the favourite
Imagine a hypothetical team proposing ways to explain a complex bill. One idea uses a timeline, another a worked receipt, and another an interactive comparison. If every option becomes a chat assistant after a brainstorming round, the team has narrowed its choices even if each proposal now reads more smoothly.
Put the options beside each other and describe what makes them different in use. A new colour scheme or a changed product name is weak evidence of a different concept. A different interaction, audience, constraint, or underlying explanation gives the team more to compare.
An assistant can help cluster the proposals, but its clusters need inspection. The same model that produced the ideas may describe superficial differences as substantial. Ask people familiar with the problem to review the actual options without seeing which source produced each one.
Record both usefulness and variety. A collection of unusual but unusable proposals has its own problem. The aim is to preserve alternatives worth testing, not maximise oddness as an end in itself.
Replace the novelty score with a search record
I would not trust a model's unsupported claim that an idea is eight out of ten for novelty. It needs a comparison set. Ask it to find related examples, then verify that those examples exist and actually resemble the proposal.
Keep a brief record of the search: terms used, sources checked, date, and the closest examples found. A web search cannot prove that nobody has had an idea before. It can reveal an obvious existing product or a recurring pattern the team should understand before building.
Changing models or increasing the number of suggestions may produce more options. It does not establish that those options are independent. The team still needs to inspect what differs.
I would judge the revised workflow by the work it leaves available for testing: several plausible approaches, their assumptions visible, and a reason to prefer one after examining the problem. The assistant can expand or edit the material. The team should retain enough of its own observations to notice when every route starts looking the same.