When AI Makes Everyone 'Creative', Creativity Itself Dies: Understanding arXiv 2605.06540 on Idea Diversity Collapse

Forum topic · 小凯 · 2026-05-08

Summary

This deep dive explains arXiv paper 2605.06540, 'Ex Ante Evaluation of AI-Induced Idea Diversity Collapse' by Azad and Baten (University of South Florida). The paper argues that current AI evaluations measure only individual utility — whether a model helps one user — while ignoring that creative outputs are consumed collectively, where value depends on scarcity. The authors formalize 'crowding' with a crowding kernel K(x,y), define an excess crowding coefficient Δ and a human-relative diversity ratio ρ, and prove that ρ ≥ 1 is the parity condition under which model adoption introduces no negative externality. Framing AI creativity as a congestible resource, they model adoption as a game where each additional user of a crowded model devalues everyone else's output. Experiments on GPT-5.4, Claude Sonnet 4.5, and Gemini 2.5 Flash across story writing, alternate-uses, and slogan tasks show all nine model-task pairs fall below parity (ρ < 1; slogans for GPT-5.4 at just 0.179). Raising sampling temperature and persona-mixture prompting partially restore diversity. The article translates the math for creators and warns of network-scale negative externalities in AI-assisted creativity.

After reading this paper, I turned off a writing assistant I had used for three years — not because it was bad, but because it was *too* good. If a hundred thousand creators use the same model for inspiration, do we end up with a hundred thousand stories that look different but share the same soul?

The Evaluation Blind Spot

Current generative AI evaluations ask one question: "Did the model help one person produce something better?" (individual utility). But creative output is consumed collectively — an idea's value depends on its scarcity in the population.

Think of a music festival: one jazz band feels fresh; ten bands playing the same standard cause fatigue. AI writing tools make it easier for everyone to produce "good" work, while making everyone's good work converge.

> Diversity collapse: when many users rely on the same generative model for inspiration, the collective diversity of human final outputs declines.

Modeling Inspiration as a Congestible Resource

Crowding Kernel

The paper defines a symmetric similarity function $K(x, y) \in [0, 1]$ measuring how much two creative outputs "crowd" each other — near 1 means crowded (similar), near 0 means distinct.

The semantic kernel maps sentence embeddings to [0, 1]:

K_{\text{sem}}(x, y) = \frac{1 + \cos(f(x), f(y))}{2}

Key Metrics

Human baseline crowding — expected crowding between two independently sampled unassisted human outputs for the same task condition $$k$$ :

\kappa^H_k = \mathbb{E}_{h, h' \sim H_k}[K_k(h, h')]

Model crowding — the same expectation over model $$m$$ 's outputs:

\kappa^A_{m,k} = \mathbb{E}_{a, a' \sim A_{m,k}}[K_k(a, a')]

Excess crowding coefficient and human-relative diversity ratio:

\Delta_{m,k} = \max\{0, \kappa^A_{m,k} - \kappa^H_k\}

\rho_{m,k} = \frac{1 - \kappa^A_{m,k}}{1 - \kappa^H_k}, \quad \kappa^H_k < 1

$\rho = 1$ means the model is as diverse as humans; $\rho > 1$ more diverse; $\rho < 1$ more homogeneous.

Proposition 1 (human-relative parity ⇔ no externality):

\Delta_{m,k} = 0 \iff \rho_{m,k} \geq 1

The Adoption Game: Why Your Convenience Devalues Others

A creator using model $$m$$ for task $$k$$ gains private benefit $$B_i$$ but suffers a uniqueness penalty that grows with the number of other users $X_{-i}$ :

C_{m,k}(X_{-i}) = \gamma_k \left(1 - \exp\{-X_{-i} \cdot \Delta_{m,k}\}\right)

where $\gamma_k$ weights the value of uniqueness in that task.

Proposition 2: adoption is rational iff $B_i > B_{\text{crit}}(X_{-i}) = \gamma_k(1 - e^{-X_{-i} \Delta_{m,k}})$ . When $\rho < 1$ , this threshold rises with exposure — the more people use the tool, the higher its hidden cost.
Proposition 3: as $X_{-i} \to \infty$ , the penalty converges to 0 if $\rho \geq 1$ , but to the full uniqueness value $\gamma_k$ if $\rho < 1$ .

Experiments: Three Frontier LLMs, All Below Parity

Models tested: GPT-5.4, Claude Sonnet 4.5, Gemini 2.5 Flash. Tasks: short stories (WritingPrompts horror/supernatural), Alternate Uses Task (AUT), and marketing slogans.

All 9 model–task combinations scored below $\rho = 1$ .

| Task | Model | $\hat{\rho}$ | $\hat{\Delta}$ | Reading | |------|-------|------|------|---------| | Slogan | GPT-5.4 | 0.179 | 0.331 | extreme convergence | | AUT | Claude 4.5 | 0.309 | 0.275 | severe convergence | | Story | GPT-5.4 | 0.372 | — | clear convergence | | Story | Gemini 2.5 | 0.446 | — | clear convergence | | Story | Claude 4.5 | 0.485 | — | clear convergence |

For slogans, GPT-5.4's diversity is only 17.9% of the human baseline. Crowding estimates stabilize with as few as 50 samples.

Interventions That Work

Higher temperature: raising sampling temperature from $$T = 1.0$$ to $$T = 1.5$$ significantly reduces crowding.
Persona-mixture prompting: prompting the model with varied Big Five personality profiles increases set-level diversity; some combinations approach or exceed the parity line.

The Uncomfortable Conclusion

What's unsettling is not that AI homogenizes creativity — that's near-common knowledge — but its mathematical inevitability: if $\rho < 1$ , each additional user imposes exponential, cumulative hidden costs on everyone. AI creative tools may carry a network negative externality: you enjoy using it while polluting the shared idea pool — the same pool you drink from.

Like a highway: the first car glides, the tenth slows, the ten-thousandth brings gridlock. The road of creativity may be congesting faster than we can see.

Paper Details

| Field | Content | |-------|---------| | Title | Ex Ante Evaluation of AI-Induced Idea Diversity Collapse | | Authors | Nafis Saami Azad, Raiyan Abdul Baten | | Affiliation | Bellini College of Artificial Intelligence, Cybersecurity, and Computing, University of South Florida | | arXiv ID | 2605.06540 | | Published | 2026-05-07 | | Categories | cs.AI, cs.GT | | Key findings | All 9 model–task pairs have $\rho < 1$ ; crowding estimates stable at 50 samples; higher temperature and persona-mixture prompting reduce crowding |