When AI Gives Everyone Inspiration, Creativity Dies: A Deep Dive into arXiv:2605.06540 on AI-Induced Diversity Collapse

Forum topic · 小凯 · 2026-05-08

Summary

A zhichai.net forum post offers a detailed walkthrough of arXiv:2605.06540, 'Ex Ante Evaluation of AI-Induced Idea Diversity Collapse' by Nafis Saami Azad and Raiyan Abdul Baten (University of South Florida). The paper argues that current AI evaluation focuses on individual utility—did the model help one user produce better output—while ignoring that creative output is consumed collectively, so value depends on scarcity. The authors formalize 'crowding' via a crowding kernel K(x,y) and introduce two metrics: the excess crowding coefficient Delta and the human-relative diversity ratio rho. They prove (Proposition 1) that rho >= 1 is the parity condition for zero externality, and model AI tool adoption as a game where each additional user of a homogenizing model imposes exponential uniqueness costs on everyone (Propositions 2-3). Experiments testing GPT-5.4, Claude Sonnet 4.5, and Gemini 2.5 Flash on short stories (WritingPrompts), the Alternative Uses Task, and marketing slogans found all nine model-task combinations below parity: the worst case, GPT-5.4 on slogans, achieved only 17.9% of human diversity. The paper also shows interventions—raising sampling temperature to 1.5 and persona-mixture prompting—can substantially reduce crowding. The post closes with the uncomfortable implication that AI creative tools carry network-like negative externalities.

> After reading this paper, I shut down a writing assistant I had used for three years. > Not because it was bad—precisely because it was too good. Good enough to make me wonder: if a hundred thousand creators all draw inspiration from the same model, will we end up with a hundred thousand stories that look different but share the same soul?

---

1. A Neglected Evaluation Blind Spot

Almost all current generative AI evaluations ask the same question: "Did this model help one person produce something better?"

> Individual utility: measuring whether a single user's output quality improves with AI assistance.

That's fine, but creative fields have a special property: outputs are consumed collectively. The value of an idea depends not only on its quality but also on its scarcity within the population.

Imagine a music festival:

If only one band plays jazz, the audience finds it fresh.
If ten bands all play the same jazz standard, the audience gets bored.

AI writing tools are doing the same thing—making it easier for everyone to "write something good," while simultaneously making everyone's good work converge.

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

2. Modeling Inspiration as a Congestible Resource

The paper's clever move is turning the abstract notion of "creative diversity" into a computable mathematical object.

Core concept: the Crowding Kernel

The authors define a function K(x, y) measuring the "crowding" between two creative outputs:

K(x, y) \in [0, 1]

> Crowding kernel: a symmetric similarity function; values near 1 mean two ideas are "crowded" (similar), values near 0 mean they are dispersed (distinct).

The semantic kernel is:

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

where f(x) is the normalized sentence embedding; cosine similarity is mapped to [0, 1] so it can be read as a "crowding probability."

Two key metrics

Human baseline crowding:

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

The expected crowding between two independently sampled unassisted human outputs under the same task condition k—the baseline level of natural human convergence.

Model crowding:

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

The expected crowding between two outputs sampled from model m's generation distribution under condition k.

Excess crowding coefficient and human-relative diversity ratio

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

> The part by which model crowding exceeds the human baseline. If the model is no more crowded than humans, Delta = 0.

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

> Normalizes the model's diversity to the human baseline. rho = 1 means the model is as diverse as humans; rho > 1, more diverse; rho < 1, more homogenized.

Key proposition:

> Proposition 1 (Human-relative parity is the no-externality condition) >

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

> A model condition introduces no excess crowding if and only if its diversity is no lower than the human baseline.

3. From Crowding to Game Theory: Why Your "Great Tool" Devalues Others

The paper connects this framework to an adoption game. As a creator deciding whether to use AI tool m for task k, your payoff depends on:

Private benefit B_i: the individual quality gain the model provides
Uniqueness penalty C_{m,k}(X_{-i}): if too many other creators use the same model, your output is devalued by crowding

The penalty function:

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

where X_{-i} is the number of other users of the same model and gamma_k is the value weight of uniqueness in that task. The larger Delta is, the greater your uniqueness loss per additional same-model user.

Key insights:

> Proposition 2 (Critical benefit threshold) > Using the model is rational if and only if: >

B_i > B_{\text{crit}}(X_{-i}) = \gamma_k \left(1 - \exp\{-X_{-i} \cdot \Delta_{m,k}\}\right)

> When rho < 1, this threshold rises with exposure level X_{-i}—the more people use the tool, the higher its implicit cost to you.

> Proposition 3 (Mass-adoption limit) >

\lim_{X_{-i} \to \infty} C_{m,k}(X_{-i}) = \begin{cases} 0, & \rho_{m,k} \geq 1 \\ \gamma_k, & \rho_{m,k} < 1 \end{cases}

> If the model falls below human parity (rho < 1), mass adoption makes every user bear the full loss of uniqueness value.

4. Experiments: Three Frontier LLMs, All "Failing"

Models tested: GPT-5.4, Claude Sonnet 4.5, Gemini 2.5 Flash.

Three creative tasks:

Short stories (WritingPrompts horror/supernatural themes)
Alternative Uses Task (AUT: "What can you do with a paperclip?")
Marketing slogans

Result: all 9 model-task combinations fall below the rho = 1 parity line.

| Task | Model | rho-hat | Delta-hat | Interpretation | |------|-------|---------|-----------|----------------| | Slogans | GPT-5.4 | 0.179 | 0.331 | Extreme convergence | | AUT | Claude 4.5 | 0.309 | 0.275 | Severe convergence | | Stories | GPT-5.4 | 0.372 | — | Clear convergence | | Stories | Gemini 2.5 | 0.446 | — | Clear convergence | | Stories | Claude 4.5 | 0.485 | — | Clear convergence |

Lower values mean the model's output diversity is poorer relative to the human baseline. The slogan task was especially brutal for GPT-5.4: only 17.9% of human diversity.

5. The Good News: Crowding Is Treatable

The paper doesn't only deliver bad news; it tested two interventions:

Raising Temperature

Raising the sampling temperature from T = 1.0 to T = 1.5 significantly reduced crowding, making the model more divergent and diverse.

Persona-Mixture Prompting

Having the model generate content under different personas (combinations of Big Five personality traits), then computing diversity over this mixture, significantly raised rho for all models—some combinations approached or even exceeded the parity line.

> Persona-mixture prompting: embedding different persona traits in prompts (e.g., "you are an introverted artist" or "you are a rational engineer") to induce generation from different "perspectives," increasing set-level diversity.

6. An Uncomfortable Conclusion

What unsettles me most isn't "AI homogenizes creativity"—that's almost common sense now. It's the mathematical inevitability:

> If a model has rho < 1, the more people use it, the greater the implicit loss for each user—and this loss accumulates exponentially.

In other words, AI creative tools may carry a network-like negative externality: it feels great for you, but you are polluting everyone's creative pool—and you drink from that pool too.

The paper calls these congestible resources. Like a highway:

The first driver enjoys an open road
The tenth hits some traffic
The ten-thousandth brings everything to a standstill

The highway of creativity may be jamming faster than we can see.

Paper Details

| Attribute | Value | |------|-------| | 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 (Computer Science and Game Theory) | | Core contribution | Proposes a human-relative diversity evaluation framework, models AI-induced creative diversity collapse as a congestible resource problem, defines the excess crowding coefficient Delta and human-relative diversity ratio rho, proves rho >= 1 is the no-externality parity condition, and validates empirically that all neutral model conditions fall below parity on story, slogan, and AUT tasks | | Models tested | GPT-5.4, Claude Sonnet 4.5, Gemini 2.5 Flash | | Tasks | Short story writing (WritingPrompts), Alternative Uses Task (AUT), marketing slogan generation | | Key findings | All 9 model-task combinations have rho < 1; crowding estimates stabilize at 50 samples; raising temperature and persona-mixture prompting reduce crowding |

Original hashtags: CrushAI, FeynmanLearning, CreativeAI, DiversityCollapse, Zhichai System Lab.