Group Selection as a Safeguard Against AI Substitution

TL;DR

The paper investigates how Generative AI influences long-term cultural evolution by modeling AI use as either a Complement or a Substitute within an agent-based, Henrich-inspired framework. It shows that AI Substitutes outperform Complements at the individual level, accelerating short-term gains but reducing variance and slowing CCE, whereas AI Complements preserve variance that can enhance long-run cultural progress, especially under strong group structure. Multi-level selection demonstrates that group-level dynamics can rescue Complement as the dominant strategy when group boundaries are strong, highlighting the importance of organizational design and policy to sustain human exploration. The findings inform strategies for AI alignment and organizational policy, advocating for diversity-preserving structures and specialized, context-specific AI systems to mitigate risks of cultural homogenization and model collapse.

Abstract

Reliance on generative AI can reduce cultural variance and diversity, especially in creative work. This reduction in variance has already led to problems in model performance, including model collapse and hallucination. In this paper, we examine the long-term consequences of AI use for human cultural evolution and the conditions under which widespread AI use may lead to "cultural collapse", a process in which reliance on AI-generated content reduces human variation and innovation and slows cumulative cultural evolution. Using an agent-based model and evolutionary game theory, we compare two types of AI use: complement and substitute. AI-complement users seek suggestions and guidance while remaining the main producers of the final output, whereas AI-substitute users provide minimal input, and rely on AI to produce most of the output. We then study how these use strategies compete and spread under evolutionary dynamics. We find that AI-substitute users prevail under individual-level selection despite the stronger reduction in cultural variance. By contrast, AI-complement users can benefit their groups by maintaining the variance needed for exploration, and can therefore be favored under cultural group selection when group boundaries are strong. Overall, our findings shed light on the long-term, population-level effects of AI adoption and inform policy and organizational strategies to mitigate these risks.

Figures (7)

• Figure 1: Comparing AI Complement and Substitute in writing tasks. Here we illustrate how people can use GenAI as a substitute or a complement in creative work. For an essay writing task on a new topic, users could use GenAI as a complement to assist their writing or ask for suggestions on the given topic, ultimately still writing the essay themselves based on the materials and structure provided by LLMs. Users could also use GenAI as a substitute and ask the LLM to generate an essay directly. Example outputs are actual answers from prompting GPT-5 with the shown questions.
• Figure 2: Effects of AI on Social Learning Outcome. Building on empirical work, we assume that AI affects the mean and variance of social learning outcomes. Greater reliance on AI leads to higher average social learning outcomes (lower $\alpha$) and reduced dispersion in outcomes (lower $\beta$). The right panel demonstrates how AI Complements (blue) and AI Substitutes (green) change the initial distribution of social learning outcomes (red).
• Figure 3: Model Iteration. The model runs in three steps iteratively: 1. Searching and learning; 2. Use of AI. 3. Selection on AI strategy. In well-mixed population, Step 3 selects for of AI strategies based on fitness. With group structure, we assume individuals interact more frequently with in-group than out-group members
• Figure 4: Cumulative outcome and adoption rate of AI Complement and AI Substitute. We run the two different conditions (with 100 repetitions) to contrast a population starting with $10 \%$ of AI Complement early adopters and one starting with $10 \%$ AI Substitute early adopters. The line plot in the upper panel shows the median skill across 100 repetitions. Error bars indicating the standard errors between runs are not visible in the figure due to their small size. We observe that even though the AI Substitute condition has a higher median at the beginning due to its immediate high efficiency, the AI Complement condition overtakes it after 18 generations of learning, due to its advantage in variance. The rate of population convergence to the full adoption of AI strategies (lower panel) also influences overall cumulative development.
• Figure 5: Selection gradient on the simplex under replicator dynamics. We plot the replicator vector field on the simplex of strategy frequencies. Payoffs $\pi_s(\mathbf{x})$ are estimated from repeated simulations of the searching-and-learning process in a well-mixed population and substituted into the replicator equation $\dot{x}_s=x_s(\pi_s(\mathbf{x})-\bar{\pi}(\mathbf{x}))$. Arrows and representative trajectories show the direction of evolutionary change from different initial compositions; filled points indicate equilibrium compositions. The background colour encodes the speed of selection. Along the edges, both AI strategies are favored in pairwise competition against no AI, but AI Help does not resist invasion by AI Substitute. Consequently, trajectories converge to the AI Substitute, implying this is the unique evolutionarily stable strategy in under standard parameters.