TL;DR: Courts and regulators are converging on the view that AI-generated content can infringe copyright when training data or outputs substantially copy protected works, while purely machine-generated output may itself lack copyright protection. The practical result: developers face liability for ingestion and output, and creators face new registration and disclosure burdens.
The Current Legal Landscape
The past eighteen months have produced a wave of rulings that clarify—unevenly—how copyright law applies to generative AI. In the United States, the Supreme Court declined to hear Thaler v. Perlmutter, leaving intact the D.C. Circuit’s holding that works lacking human authorship cannot be registered. The U.S. Copyright Office’s registration guidance now requires applicants to disclose AI-generated material and exclude it from claims, a rule that has already invalidated several registrations.
If you want to dig deeper, check out our guide on Biometric Authentication: The New Standard in Public Transit.
On the infringement side, Andersen v. Stability AI survived a motion to dismiss in part, allowing claims that the company’s models were trained on copyrighted images to proceed. Meanwhile, Getty Images v. Stability AI in the U.K. and Delaware has narrowed but not eliminated allegations of trademark and copyright misuse. The EU AI Act, effective August 2024, adds a transparency layer: general-purpose model providers must publish summaries of training data and respect opt-outs under the DSM Directive’s text-and-data-mining exceptions.
Specs and Technical Realities
Legally, the distinction between “training” and “output” matters. Training involves reproducing millions of works to compute statistical weights—courts are split on whether that is fair use. Outputs that resemble specific protected works can trigger infringement regardless of training legality. Watermarking standards such as C2PA are emerging as a compliance tool, though they are not yet legally mandated in most jurisdictions. Model cards and data provenance logs are becoming de facto evidence in litigation.
Industry Impact
Content platforms are shifting from open scraping to licensed datasets. Adobe, Shutterstock, and Getty now offer indemnified models built on cleared libraries. Startups without licensing deals face higher insurance premiums and investor caution. Creative professionals are negotiating contract clauses that prohibit AI training on their work and require disclosure of synthetic content. The net effect is a bifurcated market: compliant, licensed AI for enterprises, and high-risk, unlicensed models for experimental use.
FAQ
Q: Can I copyright AI-generated content?
A: Only if a human contributed sufficient creative expression—such as selecting, arranging, or substantially editing the output. Pure prompts alone typically do not qualify.
Q: Is training an AI model on copyrighted data always illegal?
A: No. It depends on jurisdiction and use. Some courts find training transformative and fair; others allow infringement claims to proceed. Licensing or opt-out compliance reduces risk.
Q: What should developers do to limit liability?
A: Document data provenance, honor opt-outs, use licensed or public-domain datasets where possible, add output filters, and disclose AI involvement in generated works.
Leave a Reply