The AI development giant Anthropic, which created the Claude model, has been granted ISO 42001 certification for its AI management system by the certification body Schellman. At the same time, Anthropic is being sued for copyright infringement in two separate cases, raising questions about its actual level of compliance with that standard.
ISO 42001 is ISO’s new standard for “Artificial Intelligence Management Systems,” and applies to companies that either utilize AI in their internal systems or those that include AI in their products or services. Anthropic was founded by two former employees of OpenAI and is heavily funded by Amazon.
According to an official press release issued by Anthropic:
We’re proud to be one of the first frontier AI labs to achieve this certification and hope it provides further assurance to our partners and the public of our strong commitment to AI safety.
The ISO 42001 certification builds on our existing work to develop AI safely and responsibly, including the release and recent update of our public Responsible Scaling Policy governance framework, deployment of Constitutional AI to help models operate in alignment with human values, and active research into AI safety and robustness. Schellman Compliance, LLC, an ISO certification body accredited by the ANSI National Accreditation Board, issued the certification.
In October of 2023, a large consortium of music publishers and recording industry companies sued Anthropic, and allege the AI company infringed on their copyrights to populate its large language models. Per the filing by the plaintiffs in that case:
In the process of building and operating AI models, Anthropic unlawfully copies and disseminates vast amounts of copyrighted works—including the lyrics to myriad musical compositions owned or controlled by Publishers. Publishers embrace innovation and recognize the great promise of AI when used ethically and responsibly. But Anthropic violates these principles on a systematic and widespread basis.
The court filing then lists 500 examples of copyrighted musical works that are alleged to have been absorbed by Anthropic without permission.
In a second case, authors Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson launched a class action lawsuit against Anthropic alleging “largescale theft” of their work, and the work of others, in order to populate the Claude model:
Anthropic has built a multibillion-dollar business by stealing hundreds of thousands of copyrighted books. Rather than obtaining permission and paying a fair price for the creations it exploits, Anthropic pirated them. Authors spend years conceiving, writing, and pursuing publication of their copyrighted material. The United States Constitution recognizes the fundamental principle that creators deserve compensation for their work. Yet Anthropic ignored copyright protections. An essential component of Anthropic’s business model—and its flagship “Claude” family of large language models (or “LLMs”)—is the largescale theft of copyrighted works.
ISO 42001 requires the company to implement specific controls that would prevent such infringement. One such control is related to the “Acquisition of Data” and requires the company to “determine and document details about the acquisition and selection of the data used in AI systems.” This, per ISO 42001, then includes “data rights (e.g. PII, copyright)” and “provenance of the data.”
A weakness in ISO 42001, however, is that it only requires the certified organization to develop a “statement of applicability” defining which of the ISO 42001 controls it feels are relevant and then justifying which ones it feels are not. Certification bodies like Schellman are then only expected to verify the language in the statement of applicability and not adversarially test it. In addition, auditors are typically not trained in legal matters and so cannot reliably identify when a company might exclude a control based on a faulty or fraudulent justification. This allows companies to exclude controls, such as data acquisition, based only on their opinion and a non-binding “risk assessment” that is, itself, based on a faulty and controversial definition of “risk.”
The Anthropic statement of applicability has not yet been made public, so it is not known if Anthropic excluded the copyright-related controls entirely, if Schellman failed to probe the matter, or if Schellman felt the lawsuits were, themselves, irrelevant. Furthermore, because a company is being sued for copyright infringement does not mean that it has, in fact, infringed; Schellman’s decision would likely not have been swayed by open cases where no final judgments have been issued.
If Anthropic eventually settles or is found to have infringed, however, then this will raise questions on Schellman’s certification and the validity of ISO 42001 certifications.
