Ex Ante Evaluation of AI-Induced Idea Diversity Collapse: Crowding Kernels and an Adoption Game Framework (arXiv:2605.06540)

Forum topic · 小凯 · 2026-05-08

Summary

This paper by Azad and Baten (University of South Florida, arXiv:2605.06540) introduces an ex ante framework for evaluating AI-induced diversity collapse in creative assistance. Instead of measuring individual utility after deployment, it models creative inspiration as a congestible resource: when many users draw from the same model, the collective diversity of human output erodes. The authors define a crowding kernel, an excess-crowding coefficient (Delta), and a human-relative diversity ratio (rho), provably showing that models below human parity (rho < 1) impose redundancy costs that grow with adoption. In experiments spanning GPT-5.4, Claude Sonnet 4.5, and Gemini 2.5 Flash on story writing, alternative-uses, and slogan tasks, all nine neutral model-task conditions fell below parity (lowest rho = 0.179). Estimates stabilize with ~50 samples and hold across semantic, narrative, concept, and lexical kernels. Raising temperature and persona-mixture prompting mitigate crowding. The metrics require only model generations and a human baseline, enabling diversity-impact assessment before deployment.

Ex Ante Evaluation of AI-Induced Idea Diversity Collapse

Deep-dive on arXiv:2605.06540 — *Ex Ante Evaluation of AI-Induced Idea Diversity Collapse* by Nafis Saami Azad and Raiyan Abdul Baten (Bellini College of Artificial Intelligence, Cybersecurity, and Computing, University of South Florida).

Key points

Evaluation of creative AI has traditionally focused on individual utility. The paper argues this misses a group-level externality: creative output's value depends partly on its scarcity in the population, so widespread reliance on the same model can cause AI-induced diversity collapse.
By modeling inspiration as a congestible resource, the authors show source-distribution crowding can be identified from in-distribution comparisons of model-only generations against a human baseline — enabling an ex ante (pre-deployment) protocol.
Crowding kernel: $K_{\text{sem}}(x, y) = \frac{1 + \cos(f(x), f(y))}{2}$ over normalized sentence embeddings; human ( $\kappa^H_k$ ) and model ( $\kappa^A_{m,k}$ ) crowding are expected kernel values over independently sampled pairs.
Core metrics:
Excess-crowding coefficient: $\Delta_{m,k} = \max\{0, \kappa^A_{m,k} - \kappa^H_k\}$
Human-relative diversity ratio: $\rho_{m,k} = \frac{1 - \kappa^A_{m,k}}{1 - \kappa^H_k}$ , with $\rho = 1$ the parity threshold.
Proposition 1: For $\kappa^H_k < 1$ , $\Delta_{m,k} = 0 \iff \rho_{m,k} \geq 1$ — human-relative parity exactly matches the no-externality condition.
Adoption game: a creator's net benefit is $U_i = B_i - C_{m,k}(X_{-i})$ with crowding cost $C_{m,k}(X_{-i}) = \gamma_k (1 - \exp\{-X_{-i} \cdot \Delta_{m,k}\})$ , where $\gamma_k$ weights the value of distinctiveness.
Proposition 2: adopting is rational iff $B_i > \gamma_k (1 - \exp\{-X_{-i} \cdot \Delta_{m,k}\})$ ; for sub-parity models the threshold rises monotonically with exposure.
Proposition 3: as adoption grows, cost converges to $0$ if $\rho_{m,k} \geq 1$ but to $\gamma_k$ (full loss of distinctiveness value) if $\rho_{m,k} < 1$ — a structural market failure where individually rational choices aggregate into collective diversity collapse.

Empirical evaluation

Setup: GPT-5.4, Claude Sonnet 4.5, Gemini 2.5 Flash; short stories (WritingPrompts horror/supernatural), Alternative Uses Task, marketing slogans; neutral prompts at $$T = 1.0$$ with 50 samples per condition; semantic kernel as primary, plus narrative-summary, concept-bucket, and lexical-template kernels.

Main results — all nine model–task conditions scored below parity:

| Task | Model | ρ̂ | Δ̂ | 95% CI upper bound | |------|-------|-----|-----|--------------------| | Slogan | GPT-5.4 | 0.179 | 0.331 | < 1 | | AUT | Claude 4.5 | 0.309 | 0.275 | < 1 | | Story | GPT-5.4 | 0.372 | — | < 1 | | Story | Gemini 2.5 | 0.446 | — | < 1 | | Story | Claude 4.5 | 0.485 | — | < 1 |

Robustness and stability:

Rarefaction analysis shows crowding estimates stabilize at ~50 samples — feasible for development-time benchmarking.
Sub-parity results hold across alternative kernels: narrative-summary (stories: GPT-5.4 0.509, Claude 4.5 0.594, Gemini 2.5 0.519), lexical-template (slogans), concept-bucket (AUT).

Interventions:

Raising temperature to $$T = 1.5$$ significantly reduces crowding.
Persona-mixture prompting (25-grid Big Five blend) raises $\rho$ across all models, with some combinations approaching or exceeding parity.

Practical implications and limitations

Model selection: users should prefer conditions with $\rho \geq 1$ , or at least know a model's $\rho$ .
Protocol design: temperature tuning and persona-mixture prompting are practical mitigations.
Platform responsibility: creative platforms integrating generative AI should account for crowding externalities.
Limitations: current evidence covers text-only creative tasks (image, music, code need domain-specific kernels); the human baseline relies on existing corpora; the framework measures source-distribution crowding, while actual human–AI interaction effects depend on interface and usage.

Conclusion

The paper elevates AI-induced diversity collapse from a post-hoc observation to a development-time metric. Because $\Delta$ and $\rho$ are computable from model generations alone — no expensive human-AI interaction experiments — diversity-impact assessment could become a standard pre-deployment step, much like safety evaluation.

Paper metadata

| Field | 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 |