Publishers Seek Sanctions Against OpenAI Over Chat Logs

The publishers say OpenAI represented that searching its training data and ChatGPT output logs for their copyrighted material was infeasible or unduly burdensome. They argue that later testimony showed the company had already conducted internal searches and maintained tools for measuring when models reproduced source material.

TechCrunch reports that the motion cites a database of roughly 78 million de-identified ChatGPT conversations and an internal effort called Project Giraffe. The publishers allege Project Giraffe included a Bloom-filter-based system for detecting and recording regurgitated material. They also allege OpenAI deleted or made unsearchable billions of conversations that could have been relevant to discovery.

The requested remedies include legal fees and limits on how OpenAI may use a previously produced 20-million-log sample. The court has not ruled on the latest request.

OpenAI called the accusations false and said the publishers are continuing efforts to obtain private user conversations as their case weakens. That response matters because the dispute combines two separate questions: whether relevant evidence was preserved and produced, and how courts should protect user privacy when ordering discovery from AI systems.

A March 9 court order, available through Justia, had already directed OpenAI to produce reservoirs of 78 million and 10 million logs under a de-identification protocol. The same order directed further production and discussion concerning Project Giraffe. That earlier order confirms the discovery dispute, but it does not decide the publishers' new sanctions allegations.

The practical lesson is about evidence governance rather than the underlying copyright merits. Model-evaluation datasets, output logs, filters, deletion jobs and internal measurement tools can become litigation records. Teams operating generative-AI systems need documented retention rules, reproducible searches and clear ownership of legal holds. Those controls help distinguish ordinary privacy-driven deletion from conduct that could later be challenged as evidence loss.