Cooperative AI

Impact of Early Exposure to LLM-based Chatbots
Topologies of Human–Machine Learning

Zero-sum interactions like board and computer games have attracted much interest from AI research as it surpassed humans' performance. Yet, much of human sociality consists of non-zero-sum interactions that involve cooperation and coordination. Throughout its evolutionary past, the success of the human species has largely depended on its unique cooperation abilities. Introducing AI agents as cooperation partners to social life bears immense potential but also presents the challenge of equipping AI systems with compatible capabilities to cooperate with humans. Such optimistic views go back to early thinkers like Norbert Wiener, who envisioned a symbiosis between humans and machines. Thus, this research area studies cooperative human–machine interactions—or, in short, cooperative AI.
Developing and measuring key cooperative AI concepts relies on machine behavior research. Indeed, recent behavioral studies show that dynamic reinforcement learning algorithms can establish and sustain cooperation with humans across various economic games. Across various disciplines, such as behavioral economics, human–computer interaction, and psychology, interest in settings where people and machines can cooperate is growing.

Impact of Early Exposure to LLM-based Chatbots

How does increased AI accessibility, such as the diffusion of ChatGPT, shape public sentiments toward AI?

As AI chatbots rapidly enter everyday life, an important question is how first interactions with such systems shape public sentiment about AI more broadly. This study examined the question shortly after the public release of ChatGPT, when many people still had little or no direct experience with generative chatbots.

In a preregistered experiment involving 220 US participants, individuals were randomly assigned to either have a short conversation with a generative chatbot or to imagine an interaction without actual conversation. The study assessed system-specific fear, user engagement, anthropomorphization, and potential spillover effects to other domains, including AI in medicine, recruitment and governance. Spillover effects were examined by testing whether interaction with the AI chatbot shaped broader sentiments toward AI tools across three dimensions: fear of AI tools used in specific domains, fear of AI in human-like roles (e.g., doctor, therapist; see Figure 1), and hope about AI's societal applications. It was found that the short conversation reduced fear of the chatbot and increased enjoyment, regardless of participants' prior experience. The direct interaction also fostered more realistic impressions: participants were less likely to view the system as socially sophisticated or evaluative after using it, suggesting that contact replaced speculative assumptions with grounded expectations. These exposure effects, however, were highly localized. Interaction led to lower fear of AI in familiar, concrete domains—especially medical applications—but did not reliably shift attitudes in more abstract areas such as governance, nor did it affect broader hope about AI's societal potential.

The results highlight that public sentiment toward AI is multidimensional and context-dependent. A brief encounter is sufficient to shift perceptions of the specific system involved, but does not automatically generalize to wider beliefs about AI. As AI tools continue to proliferate, these findings suggest that targeted, domain-specific engagement—rather than generic exposure—may be necessary to cultivate informed, realistic, and trusting public understanding of AI technologies.

Key Reference

Topologies of Human–Machine Learning

Do adversarial predictive systems disrupt human collective coordination, or do they spur new forms of adaptive and cultural complexity?

Contemporary AI systems are predominantly trained in a centralized manner, aggregating experience across vast populations, while interacting with humans individually as identical instantiations of the same model. This architectural asymmetry—centralized learning coupled with decentralized execution—creates a qualitatively new socio-technical topology. Humans, by contrast, coordinate through decentralized, locally constrained networks. When such centrally trained AI systems are deployed in adversarial contexts, for instance to predict, anticipate, or suppress collective human behavior, fundamental questions arise about the emergent dynamics of human–machine interaction. This raises the question whether decentralized human collectives adapt to, resist, or even outperform a centralized predictive intelligence that seeks to anticipate individual actions. Does adversarial pressure erode collective coordination, or does it catalyze new forms of collective adaptation and cultural complexity instead?

These questions are addressed using a multi-agent reinforcement learning framework that models a human–machine hybrid society (Figure 2). Decentralized agents, representing humans, are embedded in a network and rewarded for local coordination, operationalized through a graph-coloring task. A centralized predictive algorithm, representing AI, is rewarded for correctly predicting individual actions. The study systematically compares predictive architectures—personalized, shared (centrally trained and cloned), and fully global—to examine how different forms of centralization shape co-adaptive dynamics. All agents learn endogenously through deep Q-learning, allowing for open-ended behavioral adaptation.

Across conditions, decentralized agents maintain substantial levels of coordination despite adversarial pressure (Figure 3). However, competition with centralized AI drives a pronounced increase in individual behavioral complexity, measured as entropy over action sequences. Strikingly, when facing a single shared predictive model, agents converge toward highly complex yet behaviorally aligned strategies: individual actions become difficult to predict, while collective coordination remains intact. Thus, centralized prediction paradoxically induces both increased cultural complexity and reduced behavioral diversity at the group level.

These findings demonstrate that adversarial interaction between centralized AI and decentralized human collectives can generate autocurricula that accelerate cultural adaptation. Rather than collapsing under predictive control, decentralized agents evolve sophisticated coordination strategies that balance unpredictability with collective order. The work highlights both the potential risks of centralized predictive AI and the adaptive capacities of decentralized human intelligence.

Key Reference