Ambiguity Collapse by LLMs: A Taxonomy of Epistemic Risks

A team of researchers from institutions including the University of Toronto has published a significant paper on arXiv, introducing the concept of 'ambiguity collapse' in large language models (LLMs). The authors—Shira Gur-Arieh, Angelina Wang, and Sina Fazelpour—argue that when models like OpenAI's GPT-4, Anthropic's Claude, or Meta's Llama encounter value-laden, contested terms such as 'hate speech,' 'incitement,' or 'qualified,' they often produce a singular, definitive resolution. This process bypasses the essential human practices of negotiation, justification, and contestation through which meaning is socially constructed and refined.

The paper develops a three-level taxonomy of epistemic risks. At the process level, collapse forecloses opportunities for deliberation and skill development. At the output level, it distorts the concepts and reasons that guide human and AI agents. At the ecosystem level, it risks reshaping shared vocabularies and interpretive norms over time, potentially cementing a model's particular interpretation as the default. The researchers illustrate these risks with case studies involving content moderation, hiring, and constitutional AI principles.

To counter these risks, the authors sketch multi-layered mitigation principles. These span technical interventions in model training, thoughtful institutional deployment design, user interface affordances that surface ambiguity, and better management of underspecified prompts. The overarching goal is to design AI systems that can surface, preserve, and responsibly govern ambiguity rather than collapse it, recognizing ambiguity as a productive epistemic resource for society.