Keep more than one problem in view

Individual ratings and collective variation can move in opposite directions

Access to AI story suggestions improved judged creativity while making stories more alike in Doshi and Hauser’s experiment.

Short stories, UK online participants and GPT-4; not permanent creative skill or all creative work.

The generation setup changed product-idea diversity

Selected Table 3 conditions differ in semantic diversity: fresh sessions 0.452, batch generation 0.679, constraints 0.735 and personas 0.757.

Specific historical GPT-4o and student-product brief; score is 1 minus average pairwise cosine similarity, not quality. Human reference is from an earlier study.

Retaining an idea can coexist with more similar expression

AI refinement showed no detectable similarity change for the human-authored metaphor alone, but increased similarity for metaphor plus explanation.

Preprint; no equivalence established. Language background and choice were not randomized.

Promised rewards and evaluation instructions affected variation

Originality incentives reduced convergence relative to technical-quality incentives among AI-assisted writers.

Preregistered 200-person experiment reported as a preprint; everyone brainstormed without AI and ultimately received the same bonus. Adoption is an exploratory association.

Preserve different problem definitions before selecting solutions

Atlas recommends retaining distinct starting problems, examining variety across the whole collection and checking usefulness separately.

Editorial synthesis, not a tested Atlas workflow or a universal prescription. Consistency can be appropriate after the problem and solution are chosen.

Better for one person can mean less variety for everyone.

A strong idea and a strong collection of ideas are different things. Three excellent desk organisers may leave a product team with less room to explore than an organiser, a drying rail and a way to label shared food. Each proposal can be useful; the collection answers a different question: which problems have we considered?

Anil Doshi and Oliver Hauser demonstrated the tension in a 2024 experiment with 293 UK writers and 600 evaluators. Writers assigned access to GPT-4 story suggestions produced work that received higher novelty and usefulness ratings, especially when their initial creative ability was lower. Yet the AI-assisted stories were more alike. These were short stories and judged outputs, not evidence that people permanently gained or lost creative ability.

The first suggestion can quietly become the brief.

Suppose our product team asks for ideas before speaking to students. An appealing desk organiser arrives first. The next questions become how to power it, fold it or make it cheaper. The team is working hard, but inside a problem definition it never deliberately chose.

This is anchoring: a starting point shapes the alternatives considered next. In Doshi and Hauser’s experiment, assisted stories also stayed closer to the supplied AI ideas. That supports an anchoring explanation, without proving that it is the only mechanism.

A new chat resets the conversation; it does not supply a different person’s experiences. In a product-idea study published in 2026, Christian Terwiesch and colleagues found less semantic variation when GPT-4o generated one idea per fresh session than when it generated a batch. Varying personas or functional constraints increased variation. These are workflow results from a specified older model, not a ranking of current tools.

In a batch, earlier proposals remain available for comparison. Separate sessions remove that shared context. This offers a plausible explanation for the difference, not a mechanism the study isolated. A persona changes whose needs the model imagines; a functional constraint changes what a solution must satisfy. Both give the search a different starting condition.

Distance measures difference. It cannot tell us what matters.

A semantic measure maps texts to positions and compares how close they are. It can detect overlap that a count of repeated words misses. The result depends on the embedding model, the task and which text is included. It is not a universal meter of imagination.

For our product team, a glove for pigeons would be different from a desk organiser. That does not make it useful. Conversely, a small improvement to an existing product might solve a real customer problem. Product success needs evidence about users, feasibility and demand; variation alone cannot supply it.

The strategic risk is losing options before anyone evaluates them. A team may improve the average pitch while leaving an important need unexplored. That is Atlas’s interpretation of the evidence, not a measured estimate of lost innovation.

Protecting the initial idea does not protect the whole text.

Mengchen Dong and Hiromu Yakura’s July 2026 preprint makes the distinction unusually clear. Participants either wrote metaphors without AI, received AI-generated metaphor suggestions, or wrote their own metaphor and used AI to refine its explanation. AI ideation made metaphors more alike. Refining the explanation produced no statistically detectable change in similarity of the metaphors themselves—but that does not prove equal diversity.

When the researchers measured the metaphor together with its explanation, even the outputs in the refinement condition became more similar. A person can retain an original idea while its expression converges. “Human first” therefore needs a precise meaning: which part remains human-authored, and which part are we comparing?

The study included native and non-native English speakers spanning 34 native languages. Language background and the choice to interact in a native language were not randomized. This offers a wider perspective, but no causal proof that switching languages preserves creativity and no representative global estimate.

The model is only part of the creative system.

A team that rewards a polished, immediately plausible pitch may reject unusual starting points before exploring them. Better prompting cannot compensate for a selection process that keeps choosing the same kind of answer.

In an April 2026 preprint, MIT researchers Nathanael Jo and Manish Raghavan randomized 200 US/UK participants to AI access and different evaluation instructions. Everyone brainstormed without AI first. Among AI-assisted writers, being told originality would earn a bonus led to more varied final stories than emphasizing technical quality. AI-assisted writing still showed convergence; the intervention reduced it rather than demonstrating its disappearance.

Writers who pursued originality adopted less AI text verbatim. That association suggests one possible explanation, but it does not prove that copying less caused the difference. The study changed the promised reward and instructions, not actual long-term workplace incentives; everyone ultimately received the same bonus.

Keep different starting points alive long enough to test them.

Our product team needs to decide where to use AI. Atlas recommends bringing several independently observed student problems into the process before asking for solutions. Preserve their origins so that an elegant rewrite does not silently replace the need being addressed.

Then use AI to explore alternatives within each problem, expose assumptions or improve an explanation. Compare the collection before selecting a winner: are these genuinely different approaches, or new features on the same approach? Choose what to prototype using evidence about the need and the proposed solution. This is a recommendation to evaluate, not an Atlas-tested recipe.

There is a trade-off. Broad exploration costs attention, and consistency is often useful. If the task is to explain an already chosen product clearly in several channels, convergence may be desirable. If the task is to find what to build next, selecting too early can remove precisely the variation you need.

The evidence here comes from bounded writing and product tasks, including two preprints. It does not establish an inevitable future in which all culture becomes alike. The more useful question is what our creative process preserves—and what it makes easy to discard.