Yuhan Fu and Yifan Mei
Affiliation: Sun Yat-sen University
Category: Philosophy
Keywords: Philosophy of AI
Date: Friday 5th of September
Time: 15:30
Location: Gen. Henryk Dąbrowski Hall (006)
View the full session: AI
This talk will examine how conversational AI systems driven by Large Language Models (LLMs), despite having mechanisms that are fundamentally different from those of the human normative system, are becoming integrated into human social learning environments and potentially reshaping our normative cognition. We will argue that as LLMs-based conversational AI systems enter our communities as interlocutors, they create novel feedback loops in human social and moral learning that may significantly influence our normative cognition. Specifically, their conformity feature (Zhu et al., 2024)will directly influence our norm acquisition mechanism (Sripada & Stich, 2007), especially in terms of how we infer the contents of normative rules in the environment: we are probably more confident and persistent about norms acquired in the environment. At the same time, our norm execution mechanism (i.e., normative behaviours such as norm compliance and punishment behaviours) will in turn become new data or feedback for AI’s norm acquisition. These feedback loops may accelerate norm changes or reinforce preexisting norms, affecting cultural evolution in ways that demand careful examination.
Contrasting Mechanisms of Normative Acquisition We will adopt a rationalist framework (Laurence & Margolis, 2024)to understand human normative cognition: we maintain that there is an innate and richly structured normative mechanism that detects proximal normative cues in the environment, processes and represents norms, and passes information for forming normative beliefs or motivations to act in alignment with norms. The functions of the normative mechanism include norm acquisition and execution (Sripada & Stich, 2007). This rationalist understanding of normative cognition is well supported by empirical research from developmental psychology (Baillargeon et al., 2010; Hamlin, 2023; Hamlin et al., 2007; Jin & Baillargeon, 2017; Onishi & Baillargeon, 2005) In contrast, LLMs acquire normative knowledge primarily through domain-general statistical pattern extraction from vast text, architectural biases in transformer models (Vaswani et al., 2017), and alignment techniques like Reinforcement Learning from Human Feedback (Christiano et al., 2017). This creates artificial systems that can reproduce normative language but lack the real-life experiences and evolved normative systems that ground human normative cognition. They have no intrinsic drive to infer, comply with or enforce norms, nor do they possess the emotional responses that provide intuitive guidance in moral situations for humans. Despite these fundamental differences, these systems are increasingly embedded in human social environments, creating unprecedented dynamics in normative learning.
Novel Feedback Loops in Human-AI Interaction As AI systems become regular participants in human social environments, they become a significant part of our normative lives and shape our normative cognition. In particular, the human norm acquisition mechanism—responsible for detecting normative cues and inferring normative content—may be especially susceptible to AI influence. The conformity feature of LLMs ( Zhu et al., 2024 ) creates a powerful dynamic: when AI systems consistently reflect certain normative patterns or moral viewpoints, they provide precisely the kind of consistent environmental cues that our norm acquisition mechanism is designed to detect and process. Our innate conformity bias might lead us to preferentially adopt behaviours common in our community (Henrich & Gil-White, 2001), while our self-image bias causes us to selectively attend to information that reinforces existing beliefs (Mandelbaum, 2019). These biases, operating through our normative mechanisms, may be amplified when we interact with AI systems that frequently conform to user expectations and reflect dominant norms from their training data. When AI consistently frames certain moral viewpoints as common or normative to cater users, our norm acquisition mechanism may process this input as strong evidence of normative consensus, strengthening our confidence in and persistence with particular normative beliefs (Kalkstein et al., 2023). The result is a feedback loop where AI systems enhance perceived consensus around certain moral positions, and our norm execution behaviours—compliance and punishment—subsequently provide new data that reinforces these norms within AI systems. Empirical research on social media provides important context for understanding the feedback loops. For example, Robertson et al. (2024)describe how digital media create “funhouse mirror” effects that could warp our perceptions of social norms; AI systems that conform to user expectations risk creating even more distorted reflections of normative consensus. Brady and Crockett (2024) demonstrate how social media already distorts perceptions of norms by creating false impressions of what behaviours are common or approved. AI systems may further amplify these distortions, particularly as they become more integrated into social networks and communication platforms.
Possible impacts on normative performance As noted by Heyes (2024) , explicit normative cognition itself may be a “cognitive gadget” shaped by cultural evolution; AI systems are now becoming part of the environment in which these gadgets develop and operate. There might be two possible outcomes. Firstly, AI systems may accelerate normative change by amplifying feedback effects in existing learning biases. Their global reach could facilitate faster transmission of normative innovations across cultural boundaries, while their consistent availability as interaction partners may increase opportunities for normative learning. The normative influence of these systems may be particularly pronounced given their perceived authority and consistency, potentially triggering prestige bias in human learners. Alternatively, AI systems might also stagnate existing norms by providing consistent reinforcement of dominant patterns from their training data. They could crystallise a particular historical moment’s normative framework, potentially reducing normative experimentation and innovation through their conformist tendencies.
Conclusions and Future Research The conformity feature of LLMs directly interfaces with our norm acquisition mechanisms, potentially strengthening our confidence in particular normative frameworks. Meanwhile, our norm execution behaviours provide new data that further shapes AI’s representations of normative concepts. These dynamics raise important questions for future research. For example, how will the interaction between domain-specific normative mechanisms and increasingly sophisticated AI conformity shape the cultural evolution of norms? Will different demographic groups, with potentially different calibrations of their normative mechanisms, respond differently to AI normative influence? By addressing these questions through an interdisciplinary lens, we can develop a more sophisticated understanding of how these novel social actors might influence human normative psychology.