Claude analyzes and advises on Kadrey v. Meta (AI & Copyright)

Information provided to Claude: “AI Gets a Legal Gift for Thanksgiving” and Order Granting Motion to Dismiss

Based on the court documents and article provided, this case involves authors including Richard Kadrey suing Meta Platforms, Inc. over its LLaMA AI system. Specifically, the plaintiffs alleged that Meta's use of their copyrighted books to train the LLaMA models constituted copyright infringement.

The plaintiffs put forth several theories of infringement, none of which the court found viable. First, they claimed that the LLaMA models themselves are infringing derivative works because they rely on the "expressive information extracted" from the plaintiffs' books. The court rejected this, stating that the models themselves clearly do not recast or transform the books in any way that could make them derivative works.

Second, the plaintiffs argued that every output of LLaMA models infringes their books by constituting unauthorized derivative works. However, the court noted that the complaint failed to actually allege the contents of any LLaMA outputs or show how they were similar enough to the plaintiffs' works to be potentially infringing. Without specifics on infringing content, there could be no claim for vicarious infringement based on others using LLaMA models. essentially, the plaintiffs failed to demonstrate that LLaMA's outputs meet the legal standards for being derivative works.

The plaintiffs also brought claims under the DMCA related to removal of copyright management information. But again, the court found no evidence that LLaMA ever distributed copies of the books or removed the plaintiffs' ownership information.

Finally, the additional claims for unfair competition, unjust enrichment, and negligence were all found to be preempted by federal copyright law. The court ruled the plaintiffs failed to plausibly allege unlawful, unfair, or negligent conduct by Meta separate from the potential copyright issues.

In summary, while the plaintiffs took issue with Meta using their copyrighted content to train AI models without permission, they failed to demonstrate to the court's satisfaction how this actually violated their rights based on the derivative works and vicarious infringement doctrines. The ruling leaves open the possibility of revised or new cases in the future if specific evidence of similarity between outputs and source materials can be shown. But for now, this case represents a legal gift for AI developers seeking to utilize copyrighted online content as training data.

Here is an analysis of a strategy Richard Kadrey could have employed that may have better convinced the judge to rule in his favor:

The key issue is demonstrating how LLaMA's outputs are potentially infringing derivative works. So Kadrey needed to amend the complaint to include specific examples of LLaMA content that bears substantial similarity to protected expression from his novels.

For instance, he could generate sample LLaMA writings in the science fiction genre and then conduct an analysis comparing these AI outputs to his actual published writing. Any significant verbatim passages or plot sequences and characters that closely parallel his books would be evidence to allege unauthorized derivative content.

Kadrey also should have focused more on the issue of contributory infringement. Even if LLaMA itself does not directly infringe via its outputs, Meta enables users to potentially create infringing text. By training its models on full protected works without permission, Meta knowingly facilitates and induces third parties to violate Kadrey's rights. The complaint could highlight plausible statements or marketing which demonstrate that Meta actively encourages such derivative work creation through its LLaMA services. If Meta induces infringement, it would still be liable.

Additionally, Kadrey may have subpoenaed Meta for usage data and any internal assessments on the degree to which LLaMA outputs copy or transform source training materials. This discovery could uncover language directly from Meta indicating an expectation or intent for LLaMA to produce content derived from the full-text works ingested in training, including Kadrey's writings. Such documentary evidence would make contributory infringement and derivative work theories more credible.

By focusing on these aspects, Kadrey could have better pleaded infringement and convinced the judge that LLaMA does severely misappropriate copyrighted creative expression without consent rather than merely utilizing works as benign unstructured data for machine learning as Meta claimed. Tailoring the complaint and arguments in this way may have yielded a much more sympathetic ear from the court.

Here are some additional tips for Kadrey as well as recommendations on requirements the judge could impose:

Tips for Kadrey:

• Seek a preliminary injunction while the case proceeds to immediately block Meta's continued training of LLaMA using copyrighted works. Argue irreparable harm absent injunction.

• Explore opportunities to add more author plaintiffs to the case whose works are also being copied into LLaMA without permission. A class action could strengthen claim.

• Engage copyright and technology law experts to conduct independent forensic analyses of LLaMA's systems, functionality and outputs to support infringement arguments with evidence.

Recommendations for the Judge:

• Require Meta to preserve copies of all LLaMA training datasets and model iterations to allow plaintiffs to inspect materials and identify any improperly used works.

• Compel Meta to divulge all documentation about LLaMA’s system design, training methodology, and intended functionality features that may evidence inducing of infringement.

• Appoint a special master expert in AI and copyright to advise the court on the technical aspects regarding derivative work claims and examine LLaMA privately to assess infringement risks.

• Issue guidelines for what percentage of verbatim copying or abstract protected expression in outputs could potentially cross the threshold into unfixed derivative works of source training content.

By imposing such requirements, it would address the information gaps that hampered evaluation of the derivative work claims in the initial complaint. Developing formal criteria and enabling access to properly assess LLaMA's propensity for unlawful expression would aid the court in making a fully informed decision on the merits.

- by Claude