Authors’ Lawyers Inspected OpenAI’s Data—The Copyright Case Is Still Open
Lawyers representing authors in copyright litigation against OpenAI did obtain controlled access to the company’s training datasets. The review announced as a future step became an actual discovery process, but it did not produce a public inventory of the data or resolve whether OpenAI infringed the authors’ copyrights.
The dispute has since moved beyond the original inspection arrangement. The California author cases are now part of multidistrict litigation in New York, where pretrial proceedings remain pending. The secure-room episode matters because it gave qualified outsiders limited access to closely guarded evidence, not because it opened OpenAI’s training corpus to authors or the public.
What the reviewers could actually access
The arrangement covered data used to train relevant OpenAI language models and restricted access to eligible lawyers, consultants and experts. It did not give writers a copy of the corpus or let their representatives remove datasets from the controlled environment.
The review system connected a computer in a Bay Area legal office to data stored on OpenAI infrastructure. The environment limited the reviewers’ ability to transfer material elsewhere, while confidentiality rules governed anything they encountered. Searches operated against large, compressed collections rather than a simple catalogue of book titles.
That distinction changes the meaning of “seeing OpenAI’s training data.” Reviewers could examine and search material for litigation purposes, but access remained technically and legally constrained. A successful search could identify potentially relevant records for further discovery without turning the underlying collection into a portable archive.
The inspection happened—and exposed a practical conflict
The joint letter brief filed on January 17, 2025 records that the authors’ representatives had used the inspection environment six times during the first 116 days of the protocol. The same document sets out the parties’ competing accounts of whether the system provided a workable way to examine a disputed collection known as the English Colang Dataset.
The authors’ side described searches slowed by decompression and remote access, including one text query that was stopped after appearing likely to run for more than six hours. OpenAI’s section of the filing pointed to a virtual machine with 64 CPUs and 128 gigabytes of RAM and treated search technique, attendance and cooperation as the central problems. These were opposing litigation positions, not judicial findings about which party caused the difficulties.
The disagreement shows why formal access was only the beginning. A dataset may be available for inspection yet remain difficult to interrogate if its scale, storage format and search tools prevent reviewers from efficiently locating particular works. The parties consequently disputed whether the English Colang collection should remain inside the controlled system or be produced through a less restrictive method.
The inspection also did not establish what any located record would prove. Finding a book or an extract in training material could support an argument about copying, but copyright liability requires additional legal and factual analysis. Questions may include ownership, the protected elements of a work, the conduct attributable to each defendant and any fair-use defense.
The author cases moved into a broader New York proceeding
The federal transfer order entered on April 3, 2025 centralized 12 OpenAI copyright actions in the Southern District of New York for coordinated or consolidated pretrial proceedings. The order specifically identified the California cases brought by Paul Tremblay, Sarah Silverman and Michael Chabon among the actions subject to the new structure.
Centralization does not turn every plaintiff’s allegations into a single copyright claim. It allows one court to manage overlapping factual questions, discovery and pretrial disputes across cases involving authors, publishers and news organizations. Individual plaintiffs may still rely on different works, alleged uses and legal theories.
The proceeding also extends well beyond the training-data room. Discovery can encompass model-development material, licensing records, technical evidence and output logs. Training datasets and output logs should not be treated as interchangeable: the former concern material used during model development, while the latter record interactions or generated results that may answer different evidentiary questions.
What remains unresolved
The litigation was still active in the latest official case-count record available for this review. The Judicial Panel’s July 1, 2026 docket tally listed MDL No. 3143 with 18 actions pending out of 19 total actions associated with the proceeding.
That pending status is important for creators assessing the significance of the inspection. No final merits judgment identified here establishes that OpenAI’s use of the author plaintiffs’ works was lawful or unlawful. Discovery orders determine how parties can obtain and handle evidence; they do not by themselves decide infringement or fair use.
The confidentiality restrictions impose a second limit. Even if legal representatives identify relevant material, creators and the wider public may never see the underlying records unless confidentiality protections are modified or the information enters an open court filing. Access for litigation is therefore materially different from transparency about the contents and provenance of a commercial training corpus.
The lasting development is narrower than a disclosure of OpenAI’s “secret data,” but more substantial than an unfulfilled promise of access. Qualified representatives performed multiple inspections, encountered disputes over whether the system was usable, and carried those evidentiary questions into a larger federal proceeding. The copyright claims themselves remain undecided.
Also read:
Subscribe to our newsletter
Get the latest Web3, AI, and crypto news delivered straight to your inbox.