Source: Bloomberg Law

Most significant legal risks begin with model training, long before an AI system produces outputs. Large‑scale models require vast datasets that frequently are not IP‑neutral, exposing organizations to copyright risk when training on literary works, music, or photographs and to copyright protection concerns where sufficiently creative compilations of facts are involved. The fair use defense, informed by the Second Circuit’s Authors Guild v. Google framework, remains untested in the generative‑AI training context, and commercial‑scale training raises particular tension under the fourth fair use factor’s market‑harm analysis. Companies that document data sources and choose licensed content over scraped material will be better positioned in discovery.

Patent exposure can arise at the input stage: under 35 U.S.C. §271(a) a model that executes a patented method at inference may constitute direct infringement, while §271(b) may support induced infringement claims if models produce instructions users foreseeably follow; these theories lack Federal Circuit resolution in the AI context. Trade secrets pose permanent stakes because secrecy under the Defend Trade Secrets Act and state analogs is lost if confidential information becomes incorporated into model weights and globally queryable. Employee inputs into third‑party tools that allow training can destroy trade secret status, and civil liability under the DTSA can follow even for indirectly ingested misappropriated material. Mitigation requires pre‑training governance: narrow licenses, robust confidentiality, audit rights, tool vetting, technical protections, compliance monitoring, and provenance documentation, and privacy and IP diligence must proceed in parallel from the outset.