🤖 AI Summary
This paper identifies a societal content homogenization problem in human–generative AI interaction, arising from constrained user preference expression and AI retraining cycles. We develop a Bayesian interaction model between rational users and AI agents endowed with group-level preference priors, formally proving for the first time the “homogenization death spiral”: AI outputs exhibit significantly lower variance than the true user preference distribution, and retraining on AI-generated data further accelerates convergence toward homogeneity. Theoretically, we distinguish *censorship bias*—suppression of minority expressions—from *directional bias*—systematic preference shifts—and show the former severely degrades minority group utility and fairness. Through numerical simulations and equilibrium analysis, we demonstrate that interaction-friendly modeling effectively mitigates homogenization while preserving preference diversity.
📝 Abstract
When working with generative artificial intelligence (AI), users may see productivity gains, but content generated with the help of AI may not match their preferences exactly. The boost in productivity may come at the expense of users' idiosyncrasies, such as personal style and tastes, preferences we would naturally express without AI. To let users express their preferences, many AI systems let users edit their prompt (e.g., Midjourney) or allow more natural interactions (e.g., ChatGPT), and users can always review and edit the AI-generated output themselves. However, aligning a user's intentions with an AI's output can take time and may not always be worth it if the AI's first or default output "does the job." In short, users face a trade-off between AI output fidelity and communication cost. The purpose of this work is to examine the impact of this human-AI interaction on the AI-generated content we produce as a society. We propose a Bayesian model to study the societal consequences of human-AI interactions. For a given task, rational users can exchange information with the AI to align its output with their heterogeneous preferences. The AI has a knowledge of the distribution of preferences in the population and uses a Bayesian update to create the optimal output with maximal expected fidelity given the information shared by the user. Users choose the amount of information they share to maximize their utility, balancing the cost of communication with the fidelity of the output. We show that the interplay between individual users and AI may lead to societal challenges. Outputs may become more homogenized. The AI-generated output distribution has a lower variance than the users' preference distribution. And this phenomenon is exacerbated when AI-generated content is used to train the next generation of AI: we show numerically that the users' rational decisions and the AI's training process can mutually reinforce each other, leading to a homogenization "death spiral." We also study the effects of AI bias, identifying who benefits or loses when using an AI model that does not accurately reflect the population preference distribution. At the population level, the censoring type of bias (e.g., biasing against the more unique preferences) negatively impacts the population utility as a whole, especially users with uncommon preferences who rely on AI interactivity the most. On the other hand, directional biases (e.g., a slightly left-leaning AI) will influence the users' chosen output, leading to a societal bias. Nonetheless, our research also demonstrates that creating models that facilitate human-AI interactions can limit these risks and preserve the population preference diversity. A full version of this paper can be found at https://arxiv.org/abs/2309.10448.
