What Is This?
Generative AI can improve one person's output while narrowing the idea space of a whole group.
That is the tension behind the new work on AI-driven homogenization. The individual user gets a cleaner draft, more options, and faster movement. But if many people ask similar models similar questions in similar ways, their outputs can converge. The group becomes more polished and less varied.
The useful model is:
AI assistance changes two distributions:
individual quality may rise;
collective diversity may fall.
A 2026 Nature Reviews Psychology comment by Wen-Jing Yan, Lixiang Yan, Zhi-Jin Shen and colleagues argues that metacognition can help defend against this. Their available preview names three useful capacities: intellectual humility, metacognitive flexibility, and perspectival metacognition.
That sounds abstract. In practice it means: do not only ask AI for answers. Watch how it is shaping the search space.
Why Does It Matter?
Most people evaluate AI at the wrong level.
They ask:
Did this make my answer better?
That matters, but it misses the second-order effect:
Did this make everyone's answers more alike?
The problem is not that AI is useless for creativity. The opposite is often true. In a 2024 Science Advances experiment, Anil Doshi and Oliver Hauser found that access to generative AI story ideas made individual stories score as more creative, better written, and more enjoyable, especially for less creative writers. But AI-enabled stories were also more similar to each other than human-only stories.
That is the social dilemma. Each person has an incentive to use the tool because it helps their local output. But if the whole population uses the same tool in the same way, the shared idea pool can narrow.
For Jamie, this matters because Hermes, the research library, Jme-Loop, and every AI-assisted project depend on AI without wanting to become generic. The right response is not AI abstinence. It is diversity-aware AI use.
The Individual-Group Trap
AI assistance can be locally rational and globally flattening.
At the individual level, an LLM can:
- reduce blank-page cost;
- generate candidate framings;
- fill missing background;
- improve fluency;
- surface conventional options quickly.
At the group level, the same system can:
- reuse common patterns;
- overrepresent dominant cultural and linguistic priors;
- anchor many users on similar examples;
- compress unusual local knowledge into general phrasing;
- make mediocre-but-polished ideas feel finished.
This is why the main question is not whether AI helps. The question is which layer it helps and which layer it silently compresses.
What The Evidence Says
1. AI can raise individual quality while reducing collective novelty
Doshi and Hauser's story-writing experiment is the cleanest starting point. Writers who could obtain LLM-generated story ideas produced stories judged more creative, better written, and more enjoyable. The benefit was strongest for less creative writers. But the AI-assisted stories became more similar to each other.
Their interpretation is blunt: generative AI can make individual writers better off while producing a narrower collective range of novel content.
2. Homogenization shows up in creative ideation tools
Anderson, Shah, and Kreminski ran a 36-participant comparative user study of ChatGPT as a creativity support tool. Their abstract reports that ChatGPT users generated more and more detailed ideas, but different users' ideas were less semantically distinct than users of an alternative creativity support tool. ChatGPT users also felt less responsible for the ideas they generated.
This matters because homogenization is not just about final prose style. It can happen upstream, during ideation.
3. The trade-off is not inevitable
Wan and Kalman's 2026 paper is the optimistic correction. They tested whether diverse AI personas could mitigate the homogenization effect in human-AI collaborative ideation.
Their method had two phases:
- create 10 distinct AI personas and generate 300 story plots;
- give human participants either no plot, one diverse AI plot, or five diverse AI plots.
They report that diverse GenAI inputs preserved story diversity compared with a human-only baseline, with some evidence that the one-plot condition increased diversity. Their conclusion: the creativity-diversity trade-off may come from uniform deployment practices, not from an inherent limitation of generative AI.
That is the key operational insight. Similar prompts to similar models produce similar anchors. Deliberate variation can change the outcome.
4. The psychology answer is metacognition
Yan and colleagues argue that metacognition can mitigate AI-driven homogenization. The preview names three ingredients:
- Intellectual humility: knowing that the model's answer is not the full space of possible answers.
- Metacognitive flexibility: changing thinking strategies instead of staying inside the first frame.
- Perspectival metacognition: deliberately checking how the problem looks from different viewpoints.
This is the bridge from research to practice. The defence is not just "use diverse prompts." It is learning to notice when a tool is narrowing your own search process.
The Metacognition Defence
Use this as a practical protocol.
1. Generate before asking
Before asking AI for ideas, write your own rough list first.
This preserves the pre-AI baseline. It stops the model's first answer from becoming the default shape of the problem.
first: my raw map
second: AI alternatives
third: compare and recombine
2. Ask for disagreement, not polish
Polish tends to compress toward the average. Ask the model to find neglected assumptions, missing perspectives, or incompatible frames.
Bad request:
Make this better.
Better request:
Give three frames that would make this answer less conventional. For each, name the assumption it rejects.
3. Vary the generator
If the same prompt goes to the same model with the same defaults, expect similar outputs.
Variation can come from:
- different personas;
- different disciplines;
- different model families;
- different source corpora;
- different prompt constraints;
- different evaluation criteria.
Wan and Kalman's result matters because it turns this into a design variable. Diversity can be engineered into the AI-mediated process.
4. Separate ideation from selection
Do not let the same model generate, rank, polish, and finalize without an interruption.
A healthier loop is:
human baseline -> AI divergent generation -> human selection -> AI critique -> source check -> human final judgement
Each stage should have a different job. A single smooth end-to-end pass feels efficient, but it can erase weirdness before you notice it.
5. Measure sameness explicitly
For repeated work, add a diversity check. Ask:
- Does this sound like our last five outputs?
- Which source cluster is being overused?
- What perspective is missing?
- What would a non-user, competitor, critic, novice, or domain expert say?
- Which terms are generic AI residue rather than real thought?
This is why the research brief now carries a diversity check. It is a simple guard against the library converging on the same clusters because the assistant finds them easy.
Why Smart People Get This Wrong
They confuse better local output with better collective output
A better individual answer can still make the population less diverse. These are different variables.
They blame the model instead of the deployment pattern
Uniform prompts, uniform defaults, and uniform evaluation criteria create uniform outputs. The tool matters, but the workflow also matters.
They overcorrect into anti-AI purity
Avoiding AI is not the serious answer. The serious answer is preserving human priors, local context, source diversity, and adversarial frames while using AI's speed.
They ask for "creativity" as if it is one target
Creativity has multiple components: novelty, usefulness, surprise, quality, elaboration, and diversity. A model can improve one while damaging another.
How To Use This
For any AI-assisted learning, writing, strategy, or product session:
- Write the human baseline first.
- Ask AI for multiple frames, not one answer.
- Force perspective diversity.
- Keep source-backed evidence separate from synthesis.
- Check whether the result sounds like the model, the internet, or Jamie.
- Preserve odd but plausible ideas before polishing.
- Use a different pass for critique than for generation.
For Hermes specifically, the operating rule is:
AI should widen the search space before it narrows the answer.
If it only narrows, it becomes a homogenizer.
Key Terms
- AI-driven homogenization: convergence of human outputs, expressions, or ideas caused by widespread use of similar AI systems or similar AI-mediated workflows.
- Collective diversity: variation across a group of outputs, not just quality inside one output.
- Metacognition: awareness and regulation of one's own thinking process.
- Intellectual humility: recognition that one's current answer, and the model's answer, may be incomplete or biased.
- Metacognitive flexibility: ability to switch thinking strategies when a frame is too narrow.
- Perspectival metacognition: ability to inspect a problem from multiple viewpoints and notice what each viewpoint reveals or hides.
- Uniform deployment: many users using similar models, prompts, defaults, and evaluation criteria.
Recall Questions
- Why can AI improve individual creativity while reducing collective diversity?
- What is the difference between asking "did AI improve this answer?" and "did AI narrow the idea distribution?"
- Why is prompt and persona variation a design variable, not just a style choice?
- What are the three metacognitive capacities named by Yan and colleagues?
- What should be written before asking AI for ideas?
- Why should generation and selection be separated?
- What is the operating rule for Hermes-style AI use?
Best Resources To Learn More
- Start with Doshi and Hauser for the clean social-dilemma framing: individual creativity up, collective diversity down.
- Read Wan and Kalman for the design correction: diverse AI personas can preserve diversity.
- Use Yan and colleagues as the metacognition frame: the human defence is noticing how AI changes thought, not merely checking final answers.
- Use Anderson, Shah, and Kreminski for the creativity-support-tool lens: ideation tools can change the semantic distinctiveness of ideas across users.
Sources
- Wen-Jing Yan, Lixiang Yan, Zhi-Jin Shen et al., "Metacognition can mitigate AI-driven homogenization of ideas," Nature Reviews Psychology (2026). https://doi.org/10.1038/s44159-026-00587-6
- Yun Wan and Yoram M. Kalman, "Diverse AI personas can mitigate the homogenization effect in human-AI collaborative ideation," Computers in Human Behavior: Artificial Humans, 8, 100289 (2026). https://doi.org/10.1016/j.chbah.2026.100289
- Anil R. Doshi and Oliver P. Hauser, "Generative AI enhances individual creativity but reduces the collective diversity of novel content," Science Advances 10, eadn5290 (2024). https://doi.org/10.1126/sciadv.adn5290
- Barrett R. Anderson, Jash Hemant Shah and Max Kreminski, "Homogenization Effects of Large Language Models on Human Creative Ideation," Creativity and Cognition 2024 / arXiv:2402.01536. https://doi.org/10.1145/3635636.3656204
- Zhivar Sourati, Ali Shakeri Ziabari and Morteza Dehghani, "The homogenizing effect of large language models on human expression and thought," Trends in Cognitive Sciences (2026). https://doi.org/10.1016/j.tics.2026.01.003