AI Knowledge Collapse Study Provides Evidence for Market Harm in Copyright Litigation

Original Paper: Knowledge Collapse in LLMs: When Fluency Survives but Facts Fail under Recursive Synthetic Training

Authors: Figarri Keisha2,  Zekun Wu1,2,  Ze Wang1,2,  Adriano Koshiyama1,2,  Philip Treleaven2 1Holistic AI  2University College London

Original Paper: Knowledge Collapse in LLMs: When Fluency Survives but Facts Fail under Recursive Synthetic Training Authors: Figarri Keisha, Zekun Wu, Ze Wang, Adriano Koshiyama, Philip Treleaven

Executive Summary

New research from University College London demonstrates that AI models trained on their own synthetic data suffer “knowledge collapse,” a degenerative process where factual accuracy plummets while linguistic fluency remains. This finding establishes a critical, ongoing dependency on fresh, human-created data, providing plaintiff lawyers with powerful evidence to argue for market harm in copyright infringement cases.

What the Research Shows

The scarcity of high-quality, human-generated data has led AI developers to a seemingly logical solution: training new models on data generated by existing models (“synthetic data”). This paper rigorously tests that approach and finds it fundamentally flawed. Researchers subjected large language models (LLMs) to recursive training—a feedback loop where a model is trained on its own output over successive generations.

The result is a phenomenon the authors term “knowledge collapse.” Unlike total “model collapse” where the output becomes nonsensical, knowledge collapse is more insidious. The model maintains its surface-level fluency, grammar, and confident tone, but its underlying factual knowledge base degrades significantly. It begins to invent facts, forget information, and amplify its own biases, effectively becoming a “confidently wrong” system. The study proves that without a constant infusion of new, authentic, human-created content, the model’s core utility and reliability decay. This decay is not theoretical; it is a measurable and inevitable consequence of relying on a closed loop of synthetic data.

Why This Matters for Your Case

• Directly Attacks Fair Use Factor Four: This research provides empirical evidence that directly challenges the fourth fair use factor: “the effect of the use upon the potential market for or value of the copyrighted work.” By proving that LLMs require a continuous stream of human-authored content to prevent knowledge collapse, you can argue that the AI model is not a one-time transformative product but an ongoing market substitute that directly and perpetually harms the market for original works. The model’s very survival depends on the data your clients create.

• Establishes a Causal Link to Market Harm: The value of an LLM is inextricably linked to the quality and accuracy of its output. This study establishes a causal link: remove access to new, copyrighted human works, and the model’s value deteriorates. This allows you to frame the infringement not as a simple act of copying, but as the appropriation of an essential input necessary to sustain the defendant’s entire commercial enterprise, thereby supplanting the market for the original.

• Undermines the “Transformative Use” Defense: Defendants argue that their use of copyrighted data is transformative, creating something new and different. This research reframes that narrative. The “transformation” is degenerative without constant reliance on the original source material. You can argue that the LLM is not a truly independent creation but a dependent, derivative system that cannot sustain itself without the very works it infringes upon, making it less transformative and more parasitic.

Litigation Strategy

• Targeted Discovery Requests: Use this study to justify discovery requests aimed at the defendant’s training practices. Demand internal research, data logs, and communications related to model degradation, error rates over time, and the use of recursive or synthetic training data. Evidence of their own internal struggles with “model decay” or “knowledge collapse” would be invaluable in demonstrating their dependency on your client’s work.

• Deploy Expert Witnesses: Retain a computer scientist or AI researcher as an expert witness. They can use this paper’s methodology and conclusions to explain the concept of knowledge collapse to a court in clear, compelling terms. Your expert can testify that, to a reasonable degree of scientific certainty, the defendant’s model requires ongoing access to a broad corpus of human-generated works to remain commercially viable.

• Frame the Narrative for Summary Judgment and Trial: This research allows you to move beyond a simple “they copied our work” argument. Frame the defendant’s business model as inherently unsustainable without continuous infringement. Argue that this is not a case of a single past infringement but an ongoing necessity, creating a permanent and expanding harm to the creative market that fair use was never intended to protect.

Key Takeaway

This research provides the scientific foundation to argue that AI models are not self-sufficient creations but are fundamentally dependent on human-generated content for their continued accuracy and value. This dependency is a critical vulnerability in the fair use defense, allowing you to demonstrate tangible, ongoing market harm with empirical evidence.