In-depth reporting and analytical commentary on artificial intelligence regulation. No legal advice.

Developer class-action seeks to put Github AI copyright case on hold pending appellate clarification of DMCA

Context: Some of the U.S. copyright infringement lawsuits targeting AI providers like OpenAI involve § 1202 Digital Millennium Copyright Act (DMCA) claims over the alleged removal of copyright management information (CMI). In the Northern District of California, a judge dismissed those claims in a case against Github, its parent company Microsoft and its AI provider OpenAI (July 9, 2024 ai fray article), further reducing a case that had previously already been stripped of some other claims.

What’s new: On Wednesday, the class-action lawyers behind the Github case asked Judge Jon S. Tigar for permission to bring an interlocutory appeal, seeking clarification from the United States Court of Appeals for the Ninth Circuit as to whether § 1202 DMCA comes with a strict identicality requirement (PDF). The question is whether the fact that Github’s output is not identical to the asserted input disposes of any § 1202 claim as the district judge and the defendants said. The class-action lawyers would also like their case stayed pending the appellate proceedings, which is a request typically made by defendants rather than plaintiffs.

Direct impact: The defendants may want to oppose this request in order to take a consistent position that Judge Tigar rightly dismissed the DMCA claims. There are assertions in the motion for leave to appeal that defendants can reasonably dispute. But from ai fray‘s perspective, it is actually true that appellate clarification of a question that different courts (and in the Ninth Circuit’s case, arguably even the same appeals court) don’t resolve consistently would be desirable. It also makes sense to stay the Github case in the event of an appeal, given how little of it would be left short of a successful appeal.

Wider ramifications: DMCA claims and the identicality requirement are also an issue in other AI copyright cases, but other actions would presumably not be stayed, particularly not if there is a lot more substance left after adjudication a motion to dismiss than in the Github case.

Normally, appeals should be brought from final judgments. Now, the class-action lawyers at work here tend to appeal everything they can, and potentially more, and they tend to do so at all stages. But this here might be an exceptional case in which the district court would do the right thing by allowing an interlocutory appeal.

Rather than take an immediate position a few days ago, ai fray first read some of the case law referenced in the motion, as well as the motion to dismiss, the related further briefing and the order granting in part and denying in part the motion to dismiss.

The Supreme Court has not resolved the “identicality” question. There are many decisions by district courts, and they are indeed inconsistent: there are basically two schools of thought, even among district courts in the relevant circuit here (the Ninth Circuit).

That means the most important question is whether the Ninth Circuit has. The motion for leave to appeal says “the Ninth Circuit has not spoken on this point specifically,” and furthermore claims that “it has implicitly rejected an ‘identicality’ standard for the DMCA.” That characterization of Ninth Circuit law on the question is at least debatable:

In 2016, the Ninth Circuit decided Friedman v. Live Nation Merchandise. The class-action lawyers describe that decision as “holding a ”striking similarity’ between the works may give rise to a permissible inference of copying’ supporting a DMCA claim”). That interpretation is not unreasonable, but it helps to see the paragraph from which that quote is taken, as it shows that the 2016 decision was indeed about removal of CMI (a DMCA issue), though the “striking similarity” standard was adopted from some pre-DMCA case law:

“[A] ‘striking similarity’ between the works may give rise to a permissible inference of copying.” Baxter v. MCA, Inc., 812 F.2d 421, 423 (9th Cir. 1987). Here, the fact that the photographs used by Live Nation were exact copies of the images precisely as they appeared on Sony’s website and in Friedman’s book gives rise to the compelling inference that Live Nation’s photographs were directly copied from those sources. Because the only material difference in the Live Nation versions was that the CMI was missing, it was necessarily the case that the CMI had been removed on the copied version.”

There was another identicality-related decision by the Ninth Circuit that the class-action lawyers’ motion doesn’t really mention: Frost-Tsuji Architects v. Highway Inn. The dispute is mentioned, but only in the following context:

“This Court chiefly relied on Kirk Kara Corp. v. W. Stone & Metal Corp., […] (C.D. Cal. Aug. 14, 2020) in support of its conclusion that an “identicality” requirement adheres to §§1202(b)(1) and (b)(3) claims. But as noted in ADR Int’l, “[a]lthough the court in Kirk Kara held the DMCA requires identical copies, the case law it cited does not support its holding.” […]. For example, the Kirk Kara court pointed to […] Frost-Tsuji Architects v. Highway Inn, Inc., […] (D. Haw. Jan. 21, 2015), but neither case mentioned nor employed an identical copies requirement under the DMCA.”

What the motion fails to mention is that the District of Hawaii was not the final destination of that case. It went up to the Ninth Circuit. In 2017, the year after the Friedman decision, the Ninth Circuit affirmed the district court (PDF):

“We review de novo the order granting summary judgment on Frost-Tsuji’s CMI removal claim under the DMCA, id., and we affirm for the reasons stated in the district court’s order entered November 7, 2014.”

So what does that order (PDF) say? It said there was no removal of CMI when something was merely similar to works bearing CMI. To be fair, it appears that the differences between original and derivative works in Frost-Tsuji were very significant (“basing a design on FrostTsuji’s work was not the same as removing copyright management information from an original copyrighted work”). That sounds like a greater difference than in Friedman (striking similarity), but arguably what Github does with its AI coding assistant is also just to base code on pre-existing code it analyzed.

The class-action lawyers have a strong point that some line-drawing in this regard would help. They are also right that as long as some courts read an identicality requirement into the applicable statutes while others don’t, there’s disagreement among reasonable people and that, too, suggests clarification would be helpful. Whether a reasonable line would give rise to any DMCA claims against Github is, however, another question.