Relevance Score: 98/100

AI's Quest for Factual Accuracy Directly Challenges Transformative Use Defenses

Author: Yongchao Long1
AI ResearchLitigation
View Original Research on arXiv →

Original Paper: Copy-Paste to Mitigate Large Language Model Hallucinations

Authors: Yongchao Long1,2  Xian Wu3  Yingying Zhang3  Xianbin Wen1

Yuxi Zhou1,†  Shenda Hong2,†

1Department of Computer Science, Tianjin University of Technology, Tianjin, China 2National Institute of Health Data Science, Peking University, Beijing, China 3Tencent Jarvis Lab, Shenzhen, China †Corresponding author

**

Original Paper: Copy-Paste to Mitigate Large Language Model Hallucinations Authors: Yongchao Long, Xian Wu, Yingying Zhang, Xianbin Wen, Yuxi Zhou, Shenda Hong

Executive Summary

New research demonstrates that Large Language Models (LLMs) achieve factual reliability by increasing direct copying, a finding that directly contradicts and weakens the transformative use defense central to AI copyright litigation. This paper provides technical evidence that to be commercially viable, an AI’s output must be less transformative and more derivative.

What the Research Shows

AI developers face a critical business problem: LLMs often “hallucinate,” generating false or nonsensical information that undermines their reliability and commercial value. A common solution is Retrieval-Augmented Generation (RAG), where the model is given specific, trusted documents to use as context for its answers. However, even with RAG, models often ignore the provided context and invent answers. This paper, from researchers at Tianjin University of Technology and Tencent Jarvis Lab, investigates the root of this problem and proposes a direct solution.

The authors discovered a direct, inverse correlation between the degree to which an LLM copies from its provided context and its rate of hallucination. In simple terms: more copying equals fewer errors and greater factual accuracy. Based on this finding, they developed a model training method called “CopyPasteLLM.” This method explicitly trains the model to prefer high-copying responses, effectively teaching it to trust and replicate the source material rather than generating novel (and potentially incorrect) text. The results were dramatic: CopyPasteLLM achieved accuracy improvements of up to 24.5% over leading models on a key benchmark, using only a fraction of the training data. The research concludes that to be reliable, LLMs must recalibrate their behavior to favor direct copying over internal knowledge generation.

Why This Matters for Your Case

  • Directly Undermines the Transformative Use Defense: The core of the defense’s argument rests on Factor 1 of fair use—that their use of copyrighted material is “transformative.” This research provides powerful evidence to the contrary. It demonstrates that for an LLM to be accurate and commercially valuable, its use of source material must be less transformative. Copying is not an incidental byproduct; it is an engineered, necessary feature for quality.

  • Establishes Intentional, Non-Transformative Design: The creation of “CopyPasteLLM” is not an accident. It is a deliberate engineering choice to increase copying to solve a fundamental product flaw (hallucination). This allows you to argue that the defendant didn’t just happen to copy your client’s work; they likely built systems designed to copy source material faithfully to improve their product’s marketability and function.

  • Connects Copying to Commercial Value: The paper provides a clear causal link between the act of copying and the commercial value of the AI product. A more accurate, less hallucinatory model is a more valuable and marketable one. This research shows that this value is achieved through higher degrees of direct replication, strengthening your argument that the defendant is profiting directly from the expressive work of your client, not from some transformative new purpose.

Litigation Strategy

  • Targeted Discovery Requests: Use the terminology from this paper in your discovery requests. Demand all documents, communications, and source code related to “hallucination mitigation,” “contextual faithfulness,” “retrieval-augmented generation (RAG),” and any internal metrics that measure the “copying degree” or “source reliance” of model outputs. This research gives you the specific technical vocabulary to find the smoking gun.

  • Expert Witness Preparation: Your technical expert can use this paper as a foundational exhibit. They can explain to the court that the technical pursuit of accuracy is fundamentally at odds with the legal defense of transformative use. The expert can opine that any commercially viable LLM, by necessity, must implement strategies that prioritize faithful reproduction over novel generation, making its function inherently substitutive.

  • Deposition Questioning of Technical Leads: Frame questions for the defendant’s lead engineers and product managers around this trade-off.

    • “Is reducing model hallucination a primary goal for your team?”
    • “Do your methods for improving factual accuracy involve increasing the model’s reliance on provided source texts?”
    • “Would you agree that a model that faithfully reproduces factual information from a trusted source is more reliable and commercially valuable than one that invents information?” This line of questioning forces them to either admit that copying is essential to their business model or argue that their product is intentionally unreliable.

Key Takeaway

This research reframes AI’s copying from a technical bug into an essential feature for commercial viability. For plaintiff’s counsel, it provides a powerful, evidence-based argument that the core function of a reliable AI is fundamentally derivative, not transformative.