OpenAI、Microsoft 版權訴訟解封文件將訓練資料處理紀錄列作證據
Unsealed court briefs turned training data into evidence
Alter v. OpenAI and Microsoft 集體訴訟原告在解封文件中引用 OpenAI 討論 LibGen 的內部訊息,支持其指控兩家公司未經授權使用版權書籍訓練模型。
On June 15, 2022, an OpenAI corporate designee asked in Slack how concerned the company should be about mentions of LibGen across Google Docs, Slack, and GitHub. That evening, OpenAI Vice President of Research Bob McGrew wrote that it was time to remove LibGen from the company's systems and storage. Plaintiffs in the authors' copyright case now cite that exchange as part of their evidence.
The exchange appears in unsealed filings supporting a motion for partial summary judgment in Alter v. OpenAI and Microsoft, a class action brought on behalf of the Authors Guild and book copyright owners. The case is part of multidistrict litigation pending in Manhattan. The Authors Guild's account identifies the filings as Docket 1982 and Docket 1987 and says further briefing is expected, with a hearing anticipated in early 2027.
The plaintiffs allege that OpenAI and Microsoft knowingly used copyrighted books without authorization to train models. Their filings build that argument from internal messages, presentations, and later cleanup work. Training data is therefore doing two jobs in the case: it was an input to model development, and its acquisition and handling have become evidence about corporate knowledge and intent.
The provenance trail began in 2019
According to the plaintiffs' filings, Microsoft knew about OpenAI's use of LibGen by April 2019. Sam Altman and Dario Amodei presented an early version of GPT-3 to Bill Gates, Microsoft Chief Technology Officer Kevin Scott, and others, and disclosed the use of LibGen.
The filings quote Amodei describing LibGen as "a bit sketchier" as a training set. OpenAI researcher Sam McCandlish wrote that he was worried about the optics of OpenAI using copyrighted data from a "sketchy Russian website" becoming a Hacker News story. Those messages matter because they place awareness of the source inside the development record, years before the 2022 effort that the filings call "Project Clear."
The plaintiffs also use internal statements about product impact. In May 2020, OpenAI Policy Director Jack Clark wrote that better GPT systems would increasingly substitute for human labor and specifically mentioned genre fiction authors. The filing quotes him saying the work would make people unemployed and predicting that artists' concerns would likely be ignored before release.
OpenAI hired Tarun Gogineni in 2022 to improve model writing quality. The filings say he described a research mission that included having GPT models finish George R. R. Martin's A Song of Ice and Fire series. They also attribute to him the view that authors' complaints about stolen datasets and lost work were "acceptable economic disruption."
These are claims and evidence presented by the plaintiffs. The source describes a motion, further briefing, and an expected hearing. The engineering record shows how ordinary development artifacts can be assembled into a chronology: a dataset source was disclosed, employees discussed its risks, model work continued, and the company later removed related files.
Dataset governance has to survive discovery
A training pipeline can make provenance disappear operationally. Files are copied into object storage, normalized, deduplicated, tokenized, mixed with other corpora, and consumed by many runs. A final checkpoint does not provide a readable account of where each input came from or who approved it.
The governance record needs to be separate, durable, and queryable. For every corpus, record the source, acquisition method, applicable license or permission, collection date, hashes or version identifiers, approving owner, and the models or intermediate datasets that consumed it. Dataset manifests should be immutable records. Corrections can append a new state without overwriting the earlier one.
Approval also needs to be a pipeline control. A dataset marked pending or rejected should not move into preprocessing or training storage. Access policy can enforce that boundary, while build metadata ties each training run to approved manifest versions. This makes a provenance review an actual gate instead of a document that engineers can bypass.
Deletion requires dependency tracing. Removing source files does not by itself remove normalized copies, token caches, mixtures, checkpoints, evaluation samples, or logs derived from them. A removal workflow should enumerate those descendants, record what was deleted, and verify what remains. Retention and preservation decisions belong with counsel and the designated records owner, not an ad hoc cleanup in a project channel.
Internal risk discussions need the same discipline. When engineers identify a questionable source, the decision to accept or reject it should have a named owner and a linked disposition. Slack can surface the issue, but it should not be the only place where the decision exists. The record should use direct language about the source, rights, and intended use.
The unsealed material shows the practical cost of weak provenance. Years later, counsel and courts may reconstruct the system from chat messages, scattered documents, repository references, and storage changes. A maintained lineage graph gives the organization a faster, more accurate answer: which data entered the pipeline, under whose approval, and which artifacts still depend on it.
來源:DEV Community · dev.to