Conversational Alignment with Artificial Intelligence in Context

TL;DR

This paper investigates what it means for AI conversational agents to be aligned with human communicative practices, arguing that ethical alignment alone is insufficient for safe and effective dialogue. It introduces the CONTEXT-ALIGN framework, a normative set of desiderata grounded in theories of semantic content, pragmatics, common ground, discourse structure, and information flow, to evaluate how designers encode context and goals into LLMs. The authors analyze fundamental limitations of current LLM architectures—especially context window overflow and context collapse—and critique prompting and static behavioral alignment as insufficient substitutes for co-constructed context. They conclude that substantial conceptual and architectural innovations are needed to achieve genuine conversational alignment, with important safety and societal implications for high-stakes domains and everyday use. The work highlights concrete directions for improving context handling, transparency, and adaptivity to preserve the pragmatic richness of human conversation in AI systems.

Abstract

The development of sophisticated artificial intelligence (AI) conversational agents based on large language models raises important questions about the relationship between human norms, values, and practices and AI design and performance. This article explores what it means for AI agents to be conversationally aligned to human communicative norms and practices for handling context and common ground and proposes a new framework for evaluating developers' design choices. We begin by drawing on the philosophical and linguistic literature on conversational pragmatics to motivate a set of desiderata, which we call the CONTEXT-ALIGN framework, for conversational alignment with human communicative practices. We then suggest that current large language model (LLM) architectures, constraints, and affordances may impose fundamental limitations on achieving full conversational alignment.