A single jury verdict once tagged Cox Communications with an astonishing $1 billion penalty for allowing its customers to pirate more than 10,000 songs. That award was later set aside, but the underlying finding that Cox engaged in “willful contributory infringement” now sits before the U.S. Supreme Court. The stakes extend far beyond one internet service provider. If the Court sides with the nation’s largest record labels, ISPs could be forced into the role of “internet police,” terminating access for millions of Americans based on accusations alone.
Central to the controversy are peer-to-peer file-sharing protocols such as BitTorrent. Those protocols break large files into smaller “pieces” distributed across a network of users. Each participant in the “swarm” downloads and uploads pieces, with trackers coordinating the connections. This architecture makes distribution efficient but also complicates enforcement. Copyright holders often identify alleged infringers by logging IP addresses in a swarm. As studies have shown, such identification can yield a high rate of false positives since IP addresses may be associated with innocent subscribers whose networks were hijacked or spoofed.
A cable company such as Cox usually deploys automated network monitoring systems that process infringement notices. These systems match the timestamp and IP addresses with subscriber accounts, but the actual infringer is rarely identified from the account holder. The record labels contend Cox ignored thousands of infringement notices, cutting off only 32 subscribers for serial abuse while terminating over 600,000 for non-payment. The labels say this “laissez-faire attitude” maximized profits from habitual offenders. Cox counters that aggressive termination policies would risk disconnecting hospitals, universities, and households over unproven allegations.
The Supreme Court’s decision could set precedent that reverberates into artificial intelligence. Major tech firms, including Google and X, warn that if ISPs can be held liable for users’ infringement, AI platforms could face similar exposure when users generate or distribute infringing content. AI models ingest vast datasets-text, images, audio-some of which may be copyrighted. Filtering training data is technically challenging; large language models and image generators often require billions of examples, and identifying and licensing each work is practically impossible. As one legal scholar noted, “it would be impossible for an AI developer to identify and clear billions of rights claims on an individual basis.”
Recent cases highlight the tension. For instance, in Bartz v. Anthropic, the court found it fair use to train an AI on lawfully purchased books, hailing the process “spectacularly transformative.” But it rejected fair use for pirated works, regardless of transformative purpose. In Kadrey v. Meta, the judge stressed the “market harm” factor, granting summary judgment to Meta only because plaintiffs failed to present evidence of market dilution. In Thomson Reuters v. ROSS, a non-generative AI tool lost on fair use because it used copyrighted material for the same purpose as the original and created a direct market substitute.
The parallels to the ISP case are clear. Contributory infringement hinges on knowledge and facilitation-whether an ISP “knew” a user was substantially certain to infringe and failed to act, or whether an AI developer knew its training corpus contained copyrighted works and used them without authorization. Vicarious liability adds another layer: the ability to control the infringing activity, and a direct financial interest in it. For ISPs, control is limited to cutting off service; for AI companies, it could mean restricting model outputs or altering training data pipelines.
Partial solutions are provided by digital rights management technologies. ISPs can implement network-level content filtering, though this has been seen as raising privacy and net neutrality concerns. AI platforms may embed watermark detection or style filters to block outputs evoking particular artists. However, both suffer from serious engineering limitations: DRM is circumventable, and over-filtering runs the risk of suppressing lawful uses. The historical backdrop looms large. In the 1984 Betamax case, the Supreme Court held that selling a VCR did not constitute contributory infringement because the device was “capable of substantial noninfringing uses.” That standard, borrowed from patent law, shielded technologies from liability so long as they had legitimate applications.
Applying a stricter rule now could constrain not only ISPs but also emerging AI tools, cloud services, and decentralized platforms. For those who monitor the evolution of policy, it might indicate whether the Court will update the Sony standard to fit the digital ecosystem or create new rules of liability. To engineers, it’s a reminder that systems should be designed-whether for broadband delivery or AI training-that can resist legal challenges based on how they treat copyrighted material. And for the entire technology industry, the question remains: will the infrastructure of the internet and AI be built to police infringement, or to enable innovation despite it?
