Legal Proceedings and Technical Insights into AI Model Training

Anthropic, while acknowledging it offers products based on LLMs, denies the core allegations of copyright infringement^[7]. The company asserts that its use of copyrighted works falls under the protection of fair use, as defined in 17 U.S.C. § 107^[5][7]. They argue that LLMs learn patterns and relationships within data rather than storing contents, and that the responses generated by LLMs are based on a predictive process, not verbatim copying^[8]. Anthropic emphasizes that its AI models generate varied responses to similar prompts, highlighting the probabilistic nature of the technology^[8]. A key point is to show using this technology is not about expression, but rather extracting statistical information from data^[8]. Central to their defense is the claim that the training data is used to 'learn the patterns and connections between words,' similar to how humans learn^[1]. Anthropic also disputes the plaintiffs' claim that their copyrighted works were actually used in training the AI models^[7].

The plaintiffs assert that the court has subject matter jurisdiction under 28 U.S.C. §§ 1331 and 1338(a) because the action arises under the Copyright Act of 1976^[1]. They also assert personal jurisdiction over Anthropic because it has purposely conducted business in the district^[1]. Venue is claimed to be proper under 28 U.S.C. § 1400(a) and 28 U.S.C. § 1391(b)(2) due to Anthropic's infringing activities and commercialization of those activities within the district^[1].

The court set a number of deadlines in a case management order, including:

• Initial disclosures under FRCP 26 completed by October 25, 2024^[4]
• Deadline to seek leave to add new parties or amend pleadings by December 4, 2024^[4]
• Motion for class certification filed by March 6, 2025, to be heard on a 49-day track^[4]

Several key legal and factual issues have emerged as points of contention between the parties [1 1]. These include:

• Whether Anthropic’s reproduction of copyrighted works constitutes copyright infringement^[1]
• Whether Anthropic’s reproduction qualifies as fair use^[7]
• Whether the plaintiffs can demonstrate harm and are entitled to damages^[1]
• Whether Anthropic’s infringement, if any, was willful^[1]

These issues also involve technical aspects of how LLMs function, source of training data, and the nature of the AI's output^[8][7]. The court has emphasized the need for accurate briefing and representations from counsel, particularly regarding potential hazards to public health, safety, or well-being^[3].

A central aspect of the case involves the discovery of electronically stored information (ESI)^[9]. Key points regarding ESI include:

• The disclosure requirements obligate parties to disclose documents and witnesses on which they will rely^[3].
• Producing parties must search all locations with a reasonable chance of having responsive documents, including both ESI and hard copies^[3][9].
• Privilege logs must be promptly provided and sufficiently detailed to justify the privilege^[3].

To facilitate the management of ESI, a specific protocol was established, addressing aspects such as data formats, metadata fields, and redaction^[7][9]. A key component is to determine whether Anthropic used specific copyrighted materials, such as those in the Books3 dataset, for training its AI models^[5]. The court stressed candidness in these matters^[5].

Several motions and deadlines have been set forth, including a motion to dismiss^[7] and a motion for class certification^[4]. The court has emphasized that all filings must include the date and time of the hearing or conference^[3]. Initially, there was a dispute regarding the order of hearing summary judgment and class certification motions.

Judge Alsup requires plaintiff’s counsel not to engage in any class settlement discussion until after class certification^[2].

Judge Alsup also recognizes some form of pre-certification of settlement classes and recognizes there are circumstances where class members will be better served by class negotiations before certification^[2].

In any such circumstances, counsel may apply to be “interim counsel,” and ask for express authorization to negotiate on behalf of a specified putative class^[2].
The COVID-19 pandemic is no excuse to waive any local, federal, or court rules^[3].
As of August 23, 2024, full settlement discussions at any time with respect to the individual claim are permitted^[2]. Full settlement discussions as to class claims are permitted once those class claims are certified or interim counsel are appointed^[2].

The court requires both sides to promptly meet and confer and to agree on a protocol for interviewing absent putative class members^[2]. In their joint case management statement due at the outset of the case, the parties shall either describe their agreed-upon protocol or explain why no such protocol is necessary in their particular case^[2]. It has become a recurring problem in putative class actions that one or both sides may wish to interview absent putative class members regarding the merits of the case, potentially giving rise to conflict-of-interest or other ethical issues^[2]. No interviews of absent putative class members may take place unless and until the parties’ proposed protocol is approved or permission is otherwise given^[2].