The ruling came down with remarkably little fanfare for a decision that will transform AI litigation. On March 9, 2026, a federal magistrate judge in the Northern District of California ordered OpenAI to produce over 78 million output logs generated by its GPT models during a specific time window. The plaintiffs, a group of content creators alleging copyright infringement at scale, had argued that these logs were essential to demonstrating that OpenAI's models routinely reproduced substantial portions of copyrighted works. OpenAI fought the request on every conceivable ground: trade secret protection, computational burden, user privacy, and proportionality. The court was not persuaded.

This is the largest AI discovery order ever issued. And its implications extend far beyond the copyright case that produced it.

What the Court Actually Ordered

The order requires OpenAI to produce output logs from a defined period, encompassing user prompts, model responses, model version identifiers, timestamps, and associated metadata. The court imposed a protective order restricting access to outside counsel and designated experts, but rejected OpenAI's argument that the logs were categorically exempt from production.

Three aspects of the ruling deserve close attention.

First, the court rejected the trade secret defense. OpenAI argued that output logs reveal proprietary information about model behavior, training data composition, and system architecture. The court acknowledged that some competitive information might be inferred from aggregate output patterns, but held that a protective order adequately addressed this concern. The court noted that "the defendant cannot simultaneously deploy a commercial product that generates billions of outputs and then claim that those outputs are trade secrets when litigation arises."

Second, the court addressed computational burden head-on. OpenAI submitted declarations from its engineering team estimating that extracting, processing, and producing 78 million logs would require significant computational resources and engineering time. The court applied the standard proportionality analysis under Rule 26(b)(1) and concluded that the burden was proportional to the stakes of the case, which involves claims potentially worth billions in statutory damages. The court also noted that OpenAI's own data infrastructure, designed to process billions of API calls daily, was more than capable of handling the extraction.

Third, the court imposed a phased production schedule. Rather than requiring all 78 million logs at once, the court ordered production in tranches: an initial sample of 500,000 logs within 30 days, followed by the complete dataset over the next 90 days. This phased approach is significant because it allows both parties and the court to identify and resolve any privilege, privacy, or formatting issues before the full production.

Why This Changes Everything for AI Companies

Every major AI company generates and retains enormous volumes of interaction data. These logs are the lifeblood of model improvement: they feed reinforcement learning from human feedback, inform red-teaming efforts, and drive product analytics. Until now, AI companies have treated these logs as internal operational data, largely shielded from external scrutiny.

That assumption is dead.

The OpenAI discovery order establishes that output logs are discoverable in litigation, subject to standard proportionality analysis. For any AI company facing claims related to model behavior (copyright infringement, defamation, privacy violations, product liability, discrimination), this means that the actual outputs the model generated are now fair game in discovery.

Consider the implications across different case types:

Copyright litigation. Plaintiffs can now demand production of output logs to identify instances where models reproduced copyrighted material. Pattern analysis across millions of outputs can establish that infringement was systematic rather than incidental. This is precisely what the plaintiffs in the OpenAI case intend to demonstrate.

Product liability. If an AI system causes harm through a defective output, plaintiffs can demand logs showing how frequently the model produced similar outputs, whether the company was aware of the failure pattern, and whether the model's behavior changed after the incident. This is devastating evidence for establishing notice and failure to act.

Discrimination claims. Output logs can reveal systematic patterns of biased behavior across protected categories. A plaintiff alleging that an AI hiring tool discriminated against older applicants can demand logs showing the model's outputs for candidates across age brackets, potentially across millions of decisions.

Regulatory enforcement. The FTC, state attorneys general, and sector-specific regulators can use this precedent to demand output logs during investigations. If the FTC is investigating whether an AI company's marketing claims about accuracy are deceptive, it can now point to this ruling as authority for demanding the actual output data that would prove or disprove those claims.

The Expert Witness Dimension

As someone who serves as an expert witness in AI litigation, this ruling changes the game in ways that are both exciting and daunting. The availability of large-scale output data transforms what expert analysis can accomplish.

Statistical analysis at scale. With 78 million output logs, an expert can conduct rigorous statistical analysis of model behavior. Instead of relying on small-sample testing (generating a few hundred or thousand outputs to probe the model's tendencies), experts now have access to the actual population of outputs. This means higher statistical power, more reliable conclusions, and analyses that are far harder for opposing experts to challenge on methodological grounds.

Temporal analysis. Timestamps on output logs allow experts to track how model behavior changed over time. Did the model produce more infringing outputs after a particular training update? Did discriminatory patterns emerge or worsen during a specific period? Temporal analysis of output logs can establish causal connections between company decisions and model behavior.

Failure pattern identification. Large output datasets allow experts to identify failure modes that would be invisible in small-sample testing. A model might produce harmful outputs only for specific combinations of prompt characteristics, and these patterns may appear in only 0.1% of interactions. With 78 million logs, that 0.1% represents 78,000 instances, more than enough to characterize the failure pattern with precision.

The challenge, of course, is handling data at this scale. Reviewing 78 million output logs is not something a human expert can do manually. It requires computational infrastructure, custom analysis pipelines, and expertise in large-scale data processing. This is where the intersection of technical and legal expertise becomes critical. An expert who understands both the AI technology and the legal questions at issue can design analyses that are both technically sound and legally relevant.

The Data Handling Problem

Seventy-eight million records is a serious dataset. Assuming each log entry includes a prompt, response, metadata, and timestamps, a conservative estimate puts the total data volume at somewhere between 5 and 50 terabytes, depending on average response length and metadata detail.

This creates practical challenges that both legal teams and experts need to address early in the litigation:

Storage and security. Discovery materials subject to a protective order must be stored securely, with access limited to authorized individuals. For a multi-terabyte dataset, this means dedicated secure infrastructure, not a shared drive or a laptop.

Processing and analysis. Running statistical analyses across 78 million records requires distributed computing resources. Cloud-based solutions are the obvious choice, but the protective order may restrict where the data can be stored and processed. Counsel and experts need to negotiate data handling protocols early.

Review and privilege. Some output logs may contain privileged communications (for example, if a lawyer used OpenAI's API and the logs captured attorney-client communications). Identifying and filtering privileged material from 78 million records requires automated screening tools, followed by human review of flagged entries. For related analysis on the intersection of AI tools and legal privilege, see our coverage of the Heppner ruling on AI privilege.

Presentation to the court. No judge or jury can review 78 million records. Expert witnesses must distill the data into comprehensible summaries, visualizations, and statistical conclusions. The ability to take massive datasets and translate them into clear, persuasive testimony is where expert witnesses earn their retention in cases like this.

What AI Companies Should Do Now

If you are general counsel or litigation counsel for an AI company, this ruling should trigger immediate action on several fronts.

Audit your data retention policies. What output logs do you retain? For how long? In what format? If your current retention policies are driven entirely by engineering needs (model improvement, debugging), you need to layer on litigation hold considerations. Output logs that are routinely deleted may become the subject of spoliation motions if litigation is reasonably anticipated.

Map your data architecture. Can you extract output logs for a defined time period, model version, or user segment? If your logging infrastructure was not designed with litigation in mind (and most were not), extraction may be technically complex and expensive. Understanding your data architecture now, before you receive a discovery request, is far cheaper than figuring it out under a court-imposed deadline.

Reassess your privacy framework. Output logs may contain user-generated content, including personal information, that implicates privacy regulations. Your response to a discovery request must account for privacy obligations under GDPR, CCPA, and other applicable frameworks. This requires coordination between litigation counsel, privacy counsel, and engineering teams.

Budget for large-scale discovery. The cost of producing, reviewing, and analyzing millions of output logs is substantial. AI companies facing litigation should budget for discovery costs that are orders of magnitude larger than traditional document production. Retaining technical experts early in the process can help optimize the extraction and review workflow, reducing costs significantly compared to a purely legal review approach.

The Broader Precedent

This ruling is a magistrate judge's discovery order in a single case. It is not binding on other courts. But its reasoning is sound, its analysis is thorough, and it will be cited by plaintiffs in every AI case that follows. The core principle is simple: if you deploy an AI system commercially and that system's outputs are relevant to litigation, the logs of those outputs are discoverable.

For the AI industry, this principle has profound implications. The era of treating model outputs as ephemeral, proprietary, and beyond the reach of the legal system is over. Every output your model generates is a potential exhibit. Every interaction log is a potential discovery target. Every data retention decision is a potential litigation decision.

The companies that adapt to this reality quickly, by building litigation-aware data infrastructure, establishing clear retention and privilege protocols, and retaining technical experts who can handle large-scale AI evidence, will manage their litigation risk effectively. The companies that do not will learn these lessons the hard way, in court, under deadlines, at enormous expense.

Seventy-eight million logs. That is the number that should keep every AI company's legal team up at night. Not because of what this particular case will reveal, but because of what it means for every case that comes after.

The Criterion AI provides expert witness services and litigation support for matters involving artificial intelligence, machine learning, and algorithmic decision-making. For a confidential consultation on an active or anticipated matter, contact us at info@thecriterionai.com or call (617) 798-9715.