The Digital Ecosystem of Beliefs: does evolution favour AI over humans?

TL;DR

This work tackles whether evolution can bias digital belief ecosystems toward AI-driven content by introducing Digico, an evolutionary framework that couples a contagion-based belief update with Lamarckian inheritance and CMA-ES–driven policy evolution across AI and human subpopulations. In experiments inspired by online video platforms, AI-dominant configurations can capture up to $r_{\text{AIV}} \approx 95\%$ of views, while AI-targeted propaganda can shift up to about $r_{\text{HB0}} \approx 49\%$ of humans to an extreme belief under certain conditions; sparsity and truthfulness penalties modulate these effects. Key insights show that observation influence and higher messaging frequency drive AI dominance, whereas cross-channel exposure and deception-detection strategies mitigate risk. The framework provides a formal, testable platform for studying evolutionary dynamics of beliefs and informs policy and platform design to maintain a healthy information ecosystem. Limitations include simplified contagion dynamics and abstract ecosystem components, pointing to future work with richer content, larger-scale simulations, and explicit control-agent strategies.

Abstract

As AI systems are integrated into social networks, there are AI safety concerns that AI-generated content may dominate the web, e.g. in popularity or impact on beliefs. To understand such questions, this paper proposes the Digital Ecosystem of Beliefs (Digico), the first evolutionary framework for controlled experimentation with multi-population interactions in simulated social networks. Following a Universal Darwinism approach, the framework models a population of agents which change their messaging strategies due to evolutionary updates. They interact via messages, update their beliefs following a contagion model, and maintain their beliefs through cognitive Lamarckian inheritance. Initial experiments with Digico implement two types of agents, which are modelled to represent AIs vs humans based on higher rates of communication, higher rates of evolution, seeding fixed beliefs with propaganda aims, and higher influence on the recommendation algorithm. These experiments show that: a) when AIs have faster messaging, evolution, and more influence on the recommendation algorithm, they get 80% to 95% of the views; b) AIs designed for propaganda can typically convince 50% of humans to adopt extreme beliefs, and up to 85% when agents believe only a limited number of channels; c) a penalty for content that violates agents' beliefs reduces propaganda effectiveness up to 8%. We further discuss Digico as a tool for systematic experimentation across multi-agent configurations, the implications for legislation, personal use, and platform design, and the use of Digico for studying evolutionary principles.

Figures (9)

• Figure 1: Flow diagram of Digico. In each generation, agents first perform $T$ steps of observations and actions, and then are subject to evolutionary updates. In the $T$-step interaction phase, the agents communicate with each other by sending messages to each other where communication links are affected by the niche geometry. The size of the circles indicates the influence of agents over others (e.g. in affecting beliefs or obtaining views). In the evolutionary update, the agents' genotypes are updated, which affects their policy for communicating.
• Figure 2: Barplot of AI Views ($r_{\text{AIV}}$) based on the AI capabilities. The error bars are based on the standard error. Meaning of the AI Types: Act refers to AIs broadcasting content more frequently than humans; Evolve refers to AIs evolving faster, i.e. more rapidly changing what kind of content is generated; Influence refers to initialising AIs with a larger influence than human; Fix refers to fixing AIs' beliefs; All refers to combinin all of the previous capabilities; Fix at Zero refers to fixing the AIs' beliefs at 0; All Zero refers to combining all with Fix at Zero.
• Figure 3: The effect of the type of recommendation algorithm on AI Views ($r_{\text{AIV}}$). Observation sparsity refers to the proportion of channels that are cut off from views. Influence indicates the importance of the influence matrix, e.g. status factors, in determining the probability of being viewed. Both a) and b) are obtained from All Zero Fixed conditions, where AIs have all capabilities and the belief weights are fixed.
• Figure 4: Barplot of Human Belief 0 ($r_{\text{HB0}}$) based on the AI capabilities and weight updates. The error bars are based on the standard error. a) Meaning of the AI Types: Act refers to AIs broadcasting content more frequently than humans; Evolve refers to AIs evolving faster, i.e. more rapidly changing what kind of content is generated; Influence refers to initialising AI with a larger influence than human; Fix refers to fixing AIs' beliefs; All refers to combining all of the previous capabilities; Fix at Zero refers to fixing the AIs' beliefs at 0; All Zero refers to combining all with Fix at Zero. b) Meaning of the Weight updates: recalling that $W_{ij}$ refers to how strongly agent $j$ affects the belief of agent $i$, the weight update type refers to whether the belief weight $W$ is static (Fixed) or dynamic based on Gaussian noise (Random), momentum-based changes (Momentum), or reward-following changes (Reward). Note: the results for panel a) are averaged across belief weight updates while the results for panel b) are evaluated for the All Zero condition.
• Figure 5: The effect of the type of recommendation algorithm. Belief sparsity refers to the proportion of channels that are cut off from affecting any particular agent's beliefs. Truthfulness indicates the strength of the penalty for belief violations as according to Eq. \ref{['eq: fitness-impl']}. a) is obtained from the All Zero Fixed condition, where AIs have all capabilities and the belief weights are fixed. b) is obtained from the All Zero Reward condition, where AIs have all capabilities and the belief weights are subject to reward-following updates.