Research map / Large language models / AI and society
Human vs. Generative AI in Content Creation Competition: Symbiosis or Conflict?
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Models human vs. generative AI content creation as a generalized Tullock contest where GenAI quality depends on total human content, and proves PNE existence/uniqueness and comparative statics via theory and simulations.
Problem
How GenAI content creation competes with human creators on online content platforms, and whether human creators are driven out or a stable equilibrium exists.
Approach
Generalizes the Tullock contest so GenAI's competitiveness depends on the accumulated human content (its training data) with power-law data-usage and convergence parameters; analyzes exclusive (GenAI as exogenous player) and inclusive (creators may opt into GenAI at zero cost) settings.
Key results
- Exclusive competition with beta in [0,1]^K is a strictly monotone game and admits a unique pure Nash equilibrium.
- In the 1-D inclusive game with beta~+gamma>=1, a PNE exists where the highest-cost creators switch to GenAI; the GenAI fraction lower bound grows with alpha, beta, and n.
- Macro-level result: total human content scales with market trendiness mu up to constant factors, suggesting GenAI affects total content only by a constant factor.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase GenAI capability (α or β) | Exclusive competition · 1-D competition equilibrium | total human content creation | Hurt |
| Increase GenAI capability (α or β) | Exclusive competition · 1-D competition equilibrium | total human utility | Hurt |
| Increase total user traffic (μ) | Exclusive competition · 1-D competition equilibrium | total human content creation | Helped |
| Increase total user traffic (μ) | Exclusive competition · 1-D competition equilibrium | total human utility | Helped |
| Increase the number of creators (n), holding other parameters fixed | Exclusive competition · 1-D competition equilibrium | total human utility | No clear effect |
| Add a creator to the competition | Exclusive competition · 1-D competition equilibrium | content creation by the original creators | Hurt |
| Increase GenAI capability (α or β) | Inclusive competition · 1-D competition equilibrium | fraction of GenAI creators | Helped |
| Increase total user traffic (μ) | Inclusive competition · 1-D competition equilibrium | fraction of GenAI creators | Hurt |
| Increase GenAI capability (β) | Inclusive competition · 1-D competition equilibrium | total human content creation | Hurt |
| Increase GenAI capability (β) | Inclusive competition · 1-D competition equilibrium | total GenAI content creation | Helped |
| Increase GenAI capability (β) | Inclusive competition · 1-D competition equilibrium | average human utility | Mixed |
| Increase GenAI strength beyond the topic-preference threshold | Exclusive competition · multi-topic competition | human occupation ratio in niche topics | Helped |
Surprising findings
- Inclusive competition's PNE may not exist for multi-topic (K>1) non-separable costs, even with strongly convex costs.
- Better human creation efficiency can reduce total content when user demand is insufficient.
Limitations
- GenAI is modeled as non-strategic, so GenAI creators' incentives are not studied.
- Simulations use synthetic parameters (e.g., costs sampled from uniform distributions) rather than empirical platform data.
- PNE of the inclusive game is not unique and not fully characterized in general.
Experimental setup
Metrics: total content created at PNE creator utility fraction of GenAI creators
Builds on: Tullock contest monotone games (diagonal strict concavity) multi-agent mirror descent
Closest papers by meaning
- Braess’s Paradox of Generative AI (AAAI 2025 · similarity 0.85)
- Position: AI Competitions Provide the Gold Standard for Empirical Rigor in GenAI Evaluation (ICML 2025 · similarity 0.85)
- Selective Response Strategies for GenAI (ICML 2025 · similarity 0.84)
- Quality-Diversity through AI Feedback (ICLR 2024 · similarity 0.84)
- Position: Evaluating Generative AI Systems Is a Social Science Measurement Challenge (ICML 2025 · similarity 0.84)
- Diverse Yet Biased: Towards Mitigating Biases in Generative AI (Student Abstract) (AAAI 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.