Putting AI Guardrails Around Output: A Texas Two-Step Around Training Data Infringement?

The intersection of artificial intelligence (AI) technology and copyright law pits an irresistible force against an evolving and uncertain legal framework. The latest case making waves in this struggle is Concord Music Group, Inc. v. Anthropic PBC, in which Concord Music Group and other publishers alleged copyright infringement by the AI company Anthropic. One of the major issues in the case revolves around whether Anthropic’s AI models, specifically its large language models (LLMs), are generating infringing content because the results were derived from copyrighted works.

The plaintiffs sought a preliminary injunction against Anthropic. They argued that without it, publishers would suffer irreparable harm due to the potential for widespread infringement resulting from Anthropic’s AI-generated outputs. They alleged that the AI models trained on copyrighted lyrics and compositions could generate text that closely resembles or replicates protected works, enabling unauthorized use and distribution.

In its recent order adopting a settlement reached by the parties on the publisher’s motion, the court addressed a key aspect of the publishers’ request: whether Anthropic should be required to maintain its existing guardrails. In this case, “guardrails” refer to protective measures or safeguards that Anthropic has designed and implemented to prevent its AI systems from producing output that violates copyright laws. In short, these rules or filters are built into the AI to stop it from generating text that might copy or reproduce copyrighted works like song lyrics without permission.

Guardrails

The court’s order focused on Anthropic’s stated commitment to maintaining these guardrails. However, the ruling does not address the larger (and potentially more impactful) question of whether the AI’s use of copyrighted data as “training data” to learn and train its algorithms and models could itself be a legal issue.

The court noted that the parties had reached an agreement that specifically resolved the guardrails issue. Under the agreement, Anthropic was required to continue using the guardrails it had already implemented in its AI models and product offerings. The stipulation also requires that any future AI models or new products Anthropic introduces contain the guardrails, thus ensuring that they will be applied consistently.

This decision signals a significant development in how AI companies might manage the risk of copyright infringement while continuing to innovate and release new models. However, the court did not address whether collecting copyrighted material to train the AI models could be considered infringement. That is the critical unanswered question that remains at the center of the case. The order focused entirely on the use of guardrails to prevent the AI from generating infringing content, leaving unexamined the underlying issues of data collection and use.

The Unresolved Questions of Data Collection for Training Use

While the court addressed the effectiveness of guardrails, it does not resolve the fundamental issue of whether the use of copyrighted works to train AI systems could itself constitute infringement. This question remains a point of contention in many similar cases pending across the country and is a critical issue for AI developers, as well as for those generating content using AI systems.

The use of copyrighted data to train AI models raises the question of whether such training constitutes “fair use” permitted under copyright law or whether it infringes on the rights of copyright holders. Fair use is a legal doctrine that allows limited use of copyrighted material without permission from the rights holder, typically for purposes such as criticism, comment, news reporting, teaching, scholarship, or research. This issue is especially complex when it comes to large-scale data collection for training, where the data may include copyrighted works without the permission of the rights holders.

While the publishers’ motion for a preliminary injunction in the Concord Music Group case did not directly address this issue, it is clear that the debate over whether the training of AI models on copyrighted data can constitute infringement is far from settled.

Implications for the Fair Use Doctrine

The focus on guardrails rather than data collection brings into play potential implications for the fair use doctrine. The fair use exception under U.S. copyright law permits certain uses of copyrighted material without permission, typically when the use is deemed transformative, noncommercial, or beneficial to the public. In the context of AI, one could argue that if a company such as Anthropic takes significant steps to protect copyrighted material by implementing guardrails to prevent the generation of infringing content, it could strengthen the case for fair use.

Developers like Anthropic could potentially argue that its steps to mitigate infringement through the use of guardrails demonstrate a good faith effort to respect copyright laws. This proactive stance could make the fair use argument more plausible as the company is actively working to prevent the unlawful use of copyrighted content in its AI systems. By taking measures to prevent its outputs from infringing, Anthropic could bypass concerns that its collection and use of copyrighted materials for training purposes violates copyright laws. Instead, it would contend that such collection and use of copyrighted materials is part of a broader fair use framework.

On the other hand, copyright holders may argue that even with guardrails in place, the use of copyrighted data for training purposes – without explicit permission – still presents significant risks of infringement. As AI technologies evolve, the courts will likely continue to examine these issues in detail, exploring how the fair use doctrine applies in the context of AI development and use.

The Concord Music Group v. Anthropic case provides interesting insight into the ongoing legal and ethical challenges surrounding AI and copyright law. While the court’s order on the guardrails represents a step toward mitigating the risk of AI-generated output infringement, the question of whether the use of copyrighted data for training models can be considered infringement remains unanswered. As AI continues to evolve, so too will the legal landscape. The challenge lies in how companies can accelerate their innovation on the AI superhighway while navigating the evolving intellectual property traffic laws, which are gradually taking shape as cases make their way through the courts.