All articles
5 min readai / productivity

Stories got 10.7% more alike: keep AI brainstorms from converging

Give a room the same model and the same brief, and the answers start to rhyme. Anil Doshi and Oliver Hauser measured that on short stories: writers who received a plot idea from GPT-4 were rated more creative, and their stories came out 10.7% more similar to each other 1. I train teams on these tools, and the second half of that sentence is the part that costs a company something.

What the research shows.

Doshi and Hauser ran an online experiment with about 300 writers, published in Science Advances 1. One group wrote a short story unaided. The others could pull a three-sentence starting idea from GPT-4, either one idea or up to five. Independent evaluators rated the stories as more creative, better written and more enjoyable when the writer had AI access. The gain landed on the weaker writers: those who scored lowest on a creativity measure taken before the experiment produced stories rated up to 26.6% better written and 15.2% less boring. Then the authors embedded the stories and measured pairwise cosine similarity. It rose 10.7% in the one-idea condition and 8.9% in the five-idea condition. Each writer gained, and the collection narrowed.

Barrett Anderson, Jash Shah and Max Kreminski found the same shape in idea generation 2. Their 36 participants brainstormed with either ChatGPT or an Oblique Strategies deck. The set of ideas from any single participant was about as varied in both conditions, so the tool did not flatten anyone individually. Across participants, the ChatGPT group's ideas overlapped more. ChatGPT users also wrote a greater number of more detailed ideas while reporting less responsibility for them, which is the combination that makes the effect hard to notice in a room: more output, more polish, less ownership.

Vishakh Padmakumar and He He isolated where the convergence comes from 3. Their participants wrote argumentative essays in three setups: alone, with GPT-3, and with InstructGPT, the feedback-tuned version. Co-writing with InstructGPT produced a statistically reliable drop in lexical and content diversity and raised similarity between different authors' essays. GPT-3 did not. Splitting the essays by who typed what, the model's contributions carried the whole effect while the human-written spans stayed as varied as ever. Tuning a model on human preference ratings makes its prose better and its prose more alike.

The obvious workaround is to hand half the team a different model. Emily Wenger and Yoed Kenett tested that assumption and it does not hold 4. They ran standardized creativity tasks, including the Alternative Uses Test and the Divergent Association Task, across humans and a broad set of LLMs, then compared population-level spread. LLM responses resembled other LLM responses far more than humans resembled other humans, and the finding survived controls for prompt structure, verbosity, model size and vendor family. Two teams on two different frontier models land closer together than two teams of people.

The convergence tracks how you deploy the model.

Lennart Meincke, Ethan Mollick and Christian Terwiesch benchmarked 35 prompting strategies on GPT-4, generating product ideas for college students under $50 5. A pool of ideas from 100 MBA students scored 0.243 on average pairwise cosine similarity. The GPT-4 pools ran from 0.255 to 0.432 depending on the prompt, with a plain request sitting at 0.377. Chain-of-thought prompting produced the widest spread of any prompt they tested, 0.255, close to the human pool. Across these two studies, prompt design moves the spread further than vendor choice does.

Exposure design matters as much as prompt design. Joshua Ashkinaze and colleagues ran a chained experiment with more than 800 participants across 40 countries, where each person saw creative ideas from ChatGPT or from earlier participants before brainstorming 6. High AI exposure left the rated creativity of individual ideas unchanged and raised both the level and the rate of change of collective diversity. Seeding people with varied AI material widened the pool rather than narrowing it. Yun Wan and Yoram Kalman pushed the same lever: 10 structured AI personas generated 300 story plots, and writers drawing from that varied pool preserved the diversity of a human-only baseline 7.

One default prompt in one shared thread drives the convergence. A team can change that in an afternoon.

The protocol.

  1. Write before you prompt. 10 minutes, no model, 5 ideas per person on paper. In the Doshi and Hauser design the AI idea arrives before the writer has framed anything, and the frame that survives is the model's. Collect the paper ideas before opening any chat window.
  2. Split the prompt across the room. Assign each person a different angle: a constraint, a customer segment, a hostile reviewer, a budget ceiling. Ban the plain "give me 10 ideas" request, the one Meincke measured at 0.377 against 0.255 for the best variant.
  3. Make the model map the space first. Ask it to list 8 dimensions along which solutions could differ, then generate one idea per dimension. That is the chain-of-thought structure that reached human-level spread 5.
  4. Measure the overlap. Embed the final list, compute mean pairwise cosine similarity, and log the number per session. Compare it against a session where people worked on paper. The cheap version costs nothing: read the ideas aloud and count how many share a first sentence.
  5. Put a name on each idea before review. Anderson's ChatGPT users felt less responsible for what they produced 2. An owner who has to defend an idea in the next meeting reads it more critically than an anonymous bullet in a shared document.
  6. Keep one model-free track in every workshop longer than an hour. Rotate which pair runs it. Their output is your control group, and the only evidence you will have that the AI-assisted track is widening the space rather than polishing one corner of it.

For the next ideation session: 20 minutes on paper, three different prompt instructions handed out in the room, one pair working without a model, and the similarity number written at the top of the summary.

Sources.

  1. Doshi, A. R., Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28), eadn5290. doi:10.1126/sciadv.adn5290
  2. Anderson, B. R., Shah, J. H., Kreminski, M. (2024). Homogenization Effects of Large Language Models on Human Creative Ideation. Proceedings of the 16th Conference on Creativity & Cognition (C&C '24), 413-425. doi:10.1145/3635636.3656204
  3. Padmakumar, V., He, H. (2024). Does Writing with Language Models Reduce Content Diversity? ICLR 2024. arXiv:2309.05196
  4. Wenger, E., Kenett, Y. N. (2026). Large language models are homogeneously creative. PNAS Nexus, 5(3), pgag042. doi:10.1093/pnasnexus/pgag042
  5. Meincke, L., Mollick, E. R., Terwiesch, C. (2024). Prompting Diverse Ideas: Increasing AI Idea Variance. arXiv:2402.01727
  6. Ashkinaze, J., Mendelsohn, J., Qiwei, L., Budak, C., Gilbert, E. (2025). How AI Ideas Affect the Creativity, Diversity, and Evolution of Human Ideas: Evidence From a Large, Dynamic Experiment. Proceedings of the ACM Collective Intelligence Conference (CI '25). doi:10.1145/3715928.3737481
  7. Wan, Y., Kalman, Y. M. (2025). Diverse AI Personas Can Mitigate the Homogenization Effect in Human-AI Collaborative Ideation. arXiv:2504.13868

Keep reading