Knowledge Collapse
I introduced the term knowledge collapse in AI and the Problem of Knowledge Collapse, first posted to arXiv on 4 April 2024 and published in AI & Society in 2025 (DOI: 10.1007/s00146-024-02173-x).
What is knowledge collapse?
Knowledge collapse is a narrowing of the range of information that people encounter, preserve, and consider worth seeking out. Large language models can make familiar, widely represented knowledge much easier to access. Their answers may be useful, but they can give less space to specialized knowledge and less common perspectives. If many people rely on those answers instead of exploring the underlying sources, neglected knowledge may become harder for others to find and develop. Future AI systems may then draw on a narrower public record.
The paper models this as a choice facing learners: seek information through existing methods or use a cheaper AI-assisted route. Whether knowledge collapse occurs depends on the relative costs and on how much people value knowledge beyond what the AI readily provides. In the paper’s default simulation, a 20% discount on AI-generated content leaves public beliefs 2.3 times further from the truth than in the no-discount case. That figure is a result of the model, not an estimate of an observed real-world effect.
Original paper and code
Peterson, Andrew J. (2025). “AI and the Problem of Knowledge Collapse.” AI & Society 40, 3249–3269.
- arXiv record and open preprint — includes the April 2024 submission history.
- Published article and DOI.
- Replication code.
Media coverage
- The Decoder (April 2024) explains the paper and its model.
- Deepak Varuvel Dennison in Aeon (October 2025) discusses knowledge collapse in relation to local and underrepresented knowledge. The Guardian republished the essay in November 2025.
See the Press & Media page for additional coverage.