A new study finds that swapping one AI persona for a diverse set of them meaningfully cures a chatbot's tendency to sound the same every time.
Researchers frame persona diversification as a set-level design problem, testing four methods that vary along two axes: picking personas versus generating them, and spreading them evenly across a space versus pushing toward its extremes. On the Alternative Uses Task, their evolutionary persona-generation method increased response diversity by 78.8%, originality by 26.1%, flexibility by 49.5%, and overall creativity by 13.9% compared to task-only prompting, while keeping 98.5% of outputs valid. On a second benchmark, Infinity-Chat, the same method nearly doubled how far responses spread apart compared to assigning random personas. It also stacked with existing creativity-focused prompting techniques, adding another 18.6% diversity and 6.3% creativity on top.
Single-persona prompting - "answer as a marketing expert" - is already a common trick for getting more interesting output from a language model. This research suggests the trick is under-engineered: the shape of an entire set of personas, not the cleverness of any one persona, is what actually moves the needle. That matters for brainstorming tools, writing assistants, and multi-agent systems, where the goal is a spread of usable ideas rather than one polished answer.
The gains are measured on the team's own creativity benchmarks, so read "nearly doubles diversity" as an encouraging lab result, not a promise that your next AI brainstorm will stop sounding like it came from a committee.