Revisiting Rogers' Paradox in the Context of Human-AI Interaction

TL;DR

This work extends Rogers' Paradox to a human-AI setting by embedding an AI agent that learns from the entire population within a Rogers-style network. It then analyzes a suite of strategies involving humans, AI developers, and policymakers, plus a negative-feedback extension where AI interaction degrades individual learning, to assess impacts on the population's collective world model. Key findings show that cheap AI social learning by itself does not improve long-run population fitness; however, strategies such as critical social learning and selective override can elevate equilibrium understanding, while negative-feedback loops can undermine it unless mitigated by diversified learning sources. The results underscore design and policy implications—emphasizing transparency, update scheduling, and infrastructure to harness AI for expanding, not eroding, collective understanding.

Abstract

Humans learn about the world, and how to act in the world, in many ways: from individually conducting experiments to observing and reproducing others' behavior. Different learning strategies come with different costs and likelihoods of successfully learning more about the world. The choice that any one individual makes of how to learn can have an impact on the collective understanding of a whole population if people learn from each other. Alan Rogers developed simulations of a population of agents to study these network phenomena where agents could individually or socially learn amidst a dynamic, uncertain world and uncovered a confusing result: the availability of cheap social learning yielded no benefit to population fitness over individual learning. This paradox spawned decades of work trying to understand and uncover factors that foster the relative benefit of social learning that centuries of human behavior suggest exists. What happens in such network models now that humans can socially learn from AI systems that are themselves socially learning from us? We revisit Rogers' Paradox in the context of human-AI interaction to probe a simplified network of humans and AI systems learning together about an uncertain world. We propose and examine the impact of several learning strategies on the quality of the equilibrium of a society's 'collective world model'. We consider strategies that can be undertaken by various stakeholders involved in a single human-AI interaction: human, AI model builder, and society or regulators around the interaction. We then consider possible negative feedback loops that may arise from humans learning socially from AI: that learning from the AI may impact our own ability to learn about the world. We close with open directions into studying networks of human and AI systems that can be explored in enriched versions of our simulation framework.

Figures (10)

• Figure 1: Traditional Rogers' Paradox. In the leftmost panel, we depict individual learning (blue arrows). In the middle panel, we depict social learning (purple arrows). Note that social learning is delayed by one timestep. In the rightmost panel, we depict an example of humans oscillating between individual and social learning. Agents are colored by the behavior they adopt in the current timestep. Agents are considered "adapted" if their behavior matches the current state of the world, and have an increased probability of surviving to the next timestep.
• Figure 2: AI Rogers' Paradox. At each time step, humans can perform individual learning or learn from the AI system, which reverts to the population mean of the previous time step.
• Figure 3: Comparing the collective world understanding over time, in a network where agents can only learn individually and do so at a cost (left) versus a network where agents can learn individually at the same cost or learn socially for free from an AI system that socially learns from all agents in the network (right). Each network attains the same baseline expected collective world understanding: recovering the classic Rogers' Paradox finding, even with an AI node.
• Figure 4: Impact of critical social learning (thatched) from the AI over the baseline learning strategy in the presence of AI (filled). Critical social learning leads to increased population world understanding, across varying rates of world change ($u=0.01, 0.1, 0.5$); however, critical social learning is a less powerful strategy if the world is changing very rapidly.
• Figure 5: Impact of the update schedule of AI on the collective world understanding. Each dot represents the average population world understanding quality attained with the specific update schedule, in a world of a particular change rate. We consider three different rates of environment (world) change ($u$). Update rate signifies the probability that the AI will "update" (social learn) on any given timestep.