この記事は現在英語でのみご利用いただけます。
AI and Copyright: Who Owns What When AI Creates, Trains, or Copies
AI copyright governance covers three intersecting questions: can AI training on copyrighted works constitute fair use, who owns AI-generated outputs, and what are an organisation's obligations when employees use AI to create content. The US Supreme Court denied certiorari on AI authorship in March 2026. Courts are drawing clear lines. Here is what organisations need to know.
Key Takeaways
This article provides practical governance guidance verified against primary regulatory sources.
All facts and regulatory references have been verified as of May 2026.
"情報提供のみを目的としています。この記事は法律、規制、財務または専門的なアドバイスを構成するものではありません。具体的なアドバイスについては、資格を持つ専門家にご相談ください。"
AI copyright governance addresses the intellectual property questions that arise when organisations use AI systems that were trained on copyrighted works, generate outputs that may resemble protected content, or produce material whose ownership is legally uncertain. In 2026, the legal landscape is rapidly crystallising. The US Supreme Court denied certiorari in Thaler v. Perlmutter on 2 March 2026, confirming that AI-generated works without human authorship cannot be copyrighted under US law. In Bartz v. Anthropic (June 2025), the court found that training AI on legally acquired copyrighted works is transformative and supports fair use — but using pirated copies does not. The EU AI Act requires GPAI providers to publish training data summaries and respect copyright opt-outs from August 2025. For organisations, the practical governance question is not abstract — it is whether your employees' AI-generated work product is protectable, whether your AI vendors' training data was legally sourced, and what liability you face if AI outputs infringe someone else's copyright.
Training data: when is it legal?
The central copyright question in AI is whether using copyrighted works to train AI models constitutes fair use (US) or falls within text and data mining exceptions (EU). Courts are establishing clear distinctions. In Bartz v. Anthropic, the court found that Anthropic's use of legally acquired copyrighted works to train Claude was "transformative — spectacularly so" and constituted fair use. However, the court denied summary judgment for training on pirated copies — establishing that the source of training data matters as much as the purpose. In Thomson Reuters v. Ross Intelligence, the court found that using copyrighted legal headnotes to train a competing AI legal research tool was not fair use, because the AI served as a direct market substitute.
The EU Copyright Directive (Article 4) provides a text and data mining exception for research purposes and a general exception that can be overridden by rights holders through machine-readable opt-outs. The EU AI Act (Article 53, effective August 2025) requires GPAI providers to publish sufficiently detailed summaries of training data content and to comply with copyright reservation notices. The US CLEAR Act, introduced in Congress in February 2026, would establish mandatory reporting requirements for companies training AI on copyrighted works.
Output ownership: who owns what AI creates?
The US Copyright Office and courts have consistently held that copyright requires human authorship. The Supreme Court's denial of certiorari in Thaler v. Perlmutter (March 2026) confirms this principle. Material generated entirely by AI without meaningful human creative input is not copyrightable in the US. The practical implication: if an employee uses AI to generate a report, the copyrightability depends on the degree of human creative contribution — selection, arrangement, editing, and creative direction. A lightly edited AI output may not be protectable. A substantially human-directed and modified output likely is.
The position varies by jurisdiction. The UK Copyright, Designs and Patents Act (Section 9(3)) does recognise copyright in computer-generated works, attributing authorship to the person who made the arrangements necessary for creation. China's courts have granted copyright to AI-assisted works where human creative contribution was demonstrated. The EU has not yet legislated specifically on AI output ownership, though existing copyright law generally requires human authorship.
Governance framework for organisations
Organisations should implement clear policies on AI and intellectual property covering: due diligence on AI vendors' training data sources (review licence rights, terms of service, and potential exposure from using AI-generated content), employee guidelines on using AI for content creation (clarify ownership expectations, disclosure requirements, and quality review processes), output monitoring for potential infringement (scan AI outputs for potential similarity to copyrighted works before publication), and documentation of human creative contribution (maintain records of the human direction, editing, and creative input that supports copyrightability of AI-assisted outputs).
Primary sources: Norton Rose Fulbright — AI Copyright Cases Update (March 2026) | US Copyright Office — AI Training Report