A new study published in Engineering Applications of Artificial Intelligence tested more than 20 AI models, including Gemini, ChatGPT, and Llama, against more than 100 human participants on standard creativity tasks.
The finding is straightforward and uncomfortable: AI models from different companies, built on different architectures, with different training pipelines, produce ideas that cluster in the same narrow conceptual territory. Human ideas scatter. AI ideas bunch. And the gap between them is not closing.
What the Research Actually Found
The test
Researchers used divergent thinking tests. Well-established tools for measuring creative range. Tasks included brainstorming novel uses for everyday objects and generating lists of unrelated words. These tests do not measure whether a single answer is good. They measure how far a person’s or a model’s answers spread across different conceptual territories.
The individual result
On a one-to-one basis, AI models held up well. Some individual AI responses matched or edged past the average human answer in apparent originality scores.
The collective result
When researchers mapped all AI responses together for semantic similarity, the outputs clustered tightly. Human responses scattered outward across a wide conceptual space. The following shows the key differences across 5 dimensions:
| Dimension | Human Participants | AI Models (20+ tested) |
| Idea spread / conceptual range | Wide, scattered distribution | Narrow, clustered distribution |
| Individual originality score | Variable. Low to high | Consistently near average |
| Response to “be more creative” prompt | Not applicable | Marginal improvement only |
| Effect of increased randomness | Not applicable | Slight variety, lower coherence |
| Cross-model similarity | Not applicable | High. Even across different companies |
The attempted fixes
Researchers tried 2 approaches to widen AI’s creative range:
- Increasing temperature: The randomness dial on AI outputs helped slightly, but degraded coherence quickly
- Prompting models to “be more creative” or “think outside the box” nudged results marginally but did not meaningfully widen the spread.
Neither fix worked. The ceiling appears to be structural.
Why do all these different models sound the Same
The shared training data problem
The models tested came from different companies, different training pipelines, and different design philosophies. Their creative outputs still overlapped substantially. This points to something deeper than any single company’s design choices. The shared nature of the training data itself.
Every major AI model learns from an enormous but ultimately finite slice of human-written text. The internet, books, academic papers, and forums all of it reflects patterns, recurring ideas, and dominant cultural frameworks. When an AI is asked to “be creative,” it draws from that same compressed pool, shaped by what humans have already written and valued enough to publish. It can remix and recombine those patterns, but it cannot diverge beyond them the way a person with lived experience, personal stakes, and genuine surprise can.
What AI is missing
A human brainstorming session is messy because people bring in strange associations, personal memories, and random context from unrelated parts of their lives. AI has none of that. There is no Tuesday morning commute, no half-remembered conversation, no personal frustration informing the output. That absence limits how far AI ideas can stray, regardless of how clever the prompt engineering gets.
How the Major AI Models Compare on Creative Output
The following 4 AI models were among those tested in the study, all from different companies, all producing similarly clustered creative outputs across the same 5 measurable dimensions.
| AI Model | Developer | Architecture | Creative Range | Cross-Model Similarity |
| ChatGPT (GPT series) | OpenAI | Transformer LLM | Narrow/clustered | High |
| Gemini | Google DeepMind | Multimodal LLM | Narrow/clustered | High |
| Llama | Meta AI | Open-source LLM | Narrow/clustered | High |
| Claude | Anthropic | Constitutional AI | Narrow/clustered | High |
Despite 4 different developers, 4 different training approaches, and 4 different design philosophies, creative output converged across all models, confirming the problem is structural, not company-specific.
The Scale Problem That Makes This Serious
1 person using ChatGPT for a brainstorm is not a crisis. The following 3 scenarios at scale are a different matter entirely:
- Millions of people are using the same handful of AI tools: For writing, ideation, marketing copy, and product naming. All are drawn from the same underlying patterns. Gradually compresses the diversity of ideas circulating across industries
- AI-generated content duplication: Articles, essays, reports, and creative briefs produced by different people using different AI tools end up structurally similar, reducing the variety of ideas in circulation, even when each output appears original
- The ceiling effect: When people see a list of AI-generated ideas, research shows a tendency to refine and select from that list rather than extend beyond it. The AI becomes the ceiling of the ideation session, not the floor
That 3rd effect is the most insidious. The compression does not feel like compression in the moment. Each AI response feels helpful and original. The narrowing only becomes visible when you step back and look at what is not being generated. The ideas that would have existed if people had started from their own thinking rather than from a machine’s clustered output.
Why This Is an Industry-Wide Problem, Not a Single Product Failure
The study’s most significant finding is not that any 1 AI model is uncreative. It is that the convergence is consistent across models built by entirely different companies. Gemini, GPT, and Llama, trained separately, designed differently, and released by competing organisations, produce overlapping creative outputs. This means the problem is not fixable by switching to a different chatbot. It is a shared structural constraint of how large language models currently work.
Diversity-boosting techniques exist at the model architecture level, but they trade off against output quality and coherence. Researchers and developers are aware of the problem. There is no clean fix on the immediate horizon.
How to Use AI Without Letting It Shrink Your Thinking
The research does not argue that AI tools are useless for creative work. It argues that how you use them determines whether they expand or compress your thinking. The following are 4 practical approaches:
- Start without AI: Spend the first 10 minutes of any creative session generating your own ideas before opening a chatbot. Your unfiltered starting point will diverge from the AI’s clustered output in ways that reveal where your actual creative contribution lives.
- Use AI output as a launch pad, not a landing zone: Get the predictable answers out of the way fast, then push yourself to go somewhere the model clearly did not.
- Close the chat before writing your own list: After reading AI suggestions, close the conversation and write independently. Then compare. The gap between your list and the AI’s is where your creative value sits.
- Use specific, constraint-heavy prompts: Vague prompts produce the most generic clustering. Precise, unusual constraints push models toward less-traveled territory, even if imperfectly. “Give me 5 uses for a paperclip that involve water and frustration” produces more divergent output than “give me creative uses for a paperclip.”
Conclusion
The most uncomfortable finding in this research is not about AI’s limitations. It is about what happens to human thinking when AI becomes the default starting point for creative work. The tools are not making people less creative in any single interaction.
They are gradually replacing the messy, unpredictable, wide-ranging human starting point with a machine-generated cluster, and most people using them daily have not noticed the shift yet. Human creative range still outpaces these systems when measured at scale. The question is whether it will continue to, if those systems become the place where most creative work begins.
AI research, digital culture, and the ways technology is reshaping how we think and create, our newsletter covers every story worth following. Subscribe and stay ahead.
