2 Attachment(s)
Meta’s BitTorrent Uploads of ‘Pirate Library’ Data Equaled 30% of Downloads, Expert S
A lawsuit filed by several authors against Meta centers on Meta's alleged use of pirated books for AI training data and the technical details of BitTorrent which was used to obtain them. Yesterday, Meta filed a motion for summary judgment, while countering the authors' request to resolve the copyright claims in their favor. Meta's request includes new information, including the revelation that its uploads of 'pirate' library data were roughly 30% of the data it downloaded.
Over the past two years, rightsholders of all kinds have filed lawsuits against companies that develop AI models.
Most of these cases allege that AI developers used copyrighted works to train LLMs without first obtaining authorization.
Meta is among a long list of companies now being sued for this allegedly-infringing activity, including a class action lawsuit filed by authors Richard Kadrey, Sarah Silverman, and Christopher Golden. This case has a clear piracy angle, as Meta used libraries of pirated books as training material.
Meta admitted the use of these unofficial sources early on. At the same time, however, the company denied the copyright infringement allegations, noting that it would rely on a fair use defense, at least in part.
Motions for Summary Judgment
Both parties have recently submitted motions for summary judgment to the California federal court, hoping to resolve key claims before trial. The rightsholders previously argued that there is no question that Meta downloaded works from various pirate libraries, including Z-Library, Libgen, and Anna’s Archive.
The authors said that the downloading of millions of books cannot be classified as fair use. Even if that were the case, the plaintiffs say that Meta torrented a minimum of 267.4 terabytes from known pirate resources. As part of this process, Meta uploaded data to other users via BitTorrent.
Yesterday, Meta countered the authors’ position based on the argument that its use of the contested books to train AI models is fair use. Meta discusses the various fair use factors and stresses, among other things, that Meta’s alleged infringements did not cause any harm, nor can they be seen as competition for the original works.
The fair use angle, which is more detailed and nuanced than can be described here, will play a central role in most copyright related AI-lawsuits. In this case, however, Meta is also accused of uploading copyrighted content to third parties.
Meta doesn’t dispute that it uploaded data via BitTorrent but notes that there is no evidence that any data shared with others amounted to full copies of the authors’ books.
“Plaintiffs have not asserted, much less put forth undisputed facts to prove, that Meta distributed any of their works. Indeed, the most they can muster is that at least some quantity of the data Meta downloaded via torrent must have been reuploaded.”
“That does not even suffice to raise a genuine issue as to whether any distribution of Plaintiffs’ works occurred, much less allow the Court to determine as a matter of law that Meta’s copying of Plaintiffs’ works was anything but fair,” Meta adds.
In addition to opposing the authors’ motion for summary judgment, Meta also requests summary judgment, primarily focusing on the copyright infringement claim and the fair use argument.
All About BitTorrent
Both motions are accompanied by a series of expert reports, explaining the technicalities of BitTorrent to the court.
The authors’ expert, Dr. Choffnes, pointed out Meta used a script to prevent pirated data from being ‘seeded’ after downloading was complete. However, there were apparently no upload restrictions while Meta was in the process of downloading, which is referred to as the ‘leeching’ phase.
Based on the available information, Dr. Choffnes reasoned that there is “a greater than 99.99999% chance that Meta uploaded at least one piece of the plaintiffs’ works.
These findings are rebutted by an expert report from Meta’s expert, Barabara Frederiksen-Cross, who notes that the statistical model used by Dr. Choffnes is not suited to estimating the probability here. According to Cross, it “grossly inflates the probability of distribution”.
Both reports delve deep into BitTorrent specifics, including the tit-for-tat strategy, choking and unchoking, blocks and pieces, and BitTorrent’s “hole-punch” feature.
As expected, the experts arrive at opposing conclusions. The authors’ expert sees near conclusive evidence that copyrighted data was shared, while Meta’s expert sees no solid evidence to support that.
Meta Uploaded ~30% of What it Downloaded
One of the most interesting new pieces of information comes from Meta’s expert, and is based on information that was handed over during discovery last week.
According to Frederiksen-Cross, Meta used Amazon Web Services (AWS) for its torrent activity. The cost and usage data of this AWS instance shows how much content was downloaded and uploaded respectively.
“This data […] indicates that, on average, the amount of data uploaded from these AWS instances to the Internet would not have exceeded approximately 30% of the amount of data that these AWS instances downloaded from the Internet in connection with the April to July 2024 torrent download for Internet Archive, ZLib, and LibGen,” Frederiksen-Cross writes.
Meta’s expert reasons that this imbalance shows that it’s unlikely that the company uploaded the plaintiffs’ works. At the same time, however, it indicates that the company did share many terabytes of data with third parties.
It is not difficult to see how the plaintiffs could use this same data point to argue that Meta helped other people to download pirated copies of their works.
Unusable Data Blocks and Meta’s Firewall
The experts essentially differ on what can be seen as evidence of copyright infringement. Meta’s expert, for example, stresses that data is broken down into “pieces” and further into “blocks” during BitTorrent transfers. It’s argued that, even if Meta shared small blocks of data, they would be unusable to the receiver.
While this description is correct, these pieces and blocks, combined with data from other people sharing the same torrent, eventually make up the full copies of the authors’ works.
Meta’s firewall is also a point of contention. While the company specifically blocked connections from unknown parties, the authors’ expert stresses that BitTorrent’s hole-punching feature might defeat this standard firewall feature. Meta’s expert countered that this is not necessarily true.
All in all, it is clear that this case is gearing up to be much more than a straightforward copyright infringement case. The technical BitTorrent discussions alone are some of the most advanced we have ever seen in a court case.
The motions for summary judgment make it clear that both sides want the court to rule in their favor based on the currently presented evidence. While that’s certainly possible, it wouldn’t be a surprise if, given all the technicalities involved, the court sees ‘genuine disputes of material fact’ that are better suited for trial.
