EU AI Act — transparency and general-purpose obligations
2 August 2025 and 2 August 2026 · European Union
General-purpose AI obligations began applying on 2 August 2025, including a copyright policy and public training-content summary for covered providers. Commission enforcement powers apply from 2 August 2026. Models placed on the market before 2 August 2025 have a transition period to 2 August 2027. Article 50 addresses separate transparency duties.
Determine which provider, model, and system obligations apply before deciding what must be disclosed.
Practical record to retain: A source inventory, relevant rights-reservation records, the copyright policy, and the basis for the applicable training summary.
Read the source
Bartz v. Anthropic
June 2025 ruling; settlement approved July 2026 · Northern District of California
The June 2025 ruling treated the training use at issue as fair use while distinguishing acquisition of pirated copies for a central library. Final approval of a $1.5 billion settlement followed in July 2026.
The result is specific to the case. It does not establish a universal exemption for AI training.
Practical record to retain: Source acquisition records, licenses or other asserted legal bases, and the intended training use.
Read the source
California AB 2013 — Generative AI Training Data Transparency Act
Disclosure deadline 1 January 2026 · California
AB 2013 requires covered developers making qualifying generative AI systems or services available to Californians to publish a high-level training-data summary. The statute addresses systems released from 1 January 2022 and contains scope definitions and exemptions. Its disclosure fields include sources, intellectual-property information, personal information, and synthetic-data use.
The obligation attaches to covered developers and releases; it is not a blanket disclosure duty for every dataset supplier.
Practical record to retain: The dataset inventory and supporting facts needed for the applicable statutory summary, including synthetic-data use.
Read the source
hiQ v. LinkedIn and Meta v. Bright Data
April 2022 and January 2024 decisions · United States
The Ninth Circuit’s hiQ decision addressed the Computer Fraud and Abuse Act in a preliminary-injunction dispute involving public LinkedIn profiles. Meta v. Bright Data addressed the scope of Meta’s contractual terms for the public-data collection at issue.
These decisions address different claims and facts. Public access does not automatically resolve copyright, privacy, contracts, or training rights.
Practical record to retain: Source access conditions, authorization, collection times, applicable terms, and the rights or legal basis relied on.
Read the source · Read the Bright Data order
Thomson Reuters v. Ross Intelligence
11 February 2025 decision · District of Delaware
The district court rejected fair use for the copying of protected Westlaw headnotes used to develop the competing legal research tool at issue. The decision considered purpose and market effects in its fact-specific analysis.
An absence of a licensing market is not by itself permission to copy. Assess the material, use, and applicable legal basis.
Practical record to retain: The source and intended use, relevant license terms, and the evidence supporting any legal exception relied on.
Read the source
25 June 2025 decision · Northern District of California
The court granted Meta partial summary judgment on the training claim presented by thirteen authors. It emphasized limitations in the plaintiffs’ market-harm evidence and expressly limited what the ruling established.
This ruling was not a general declaration that unlicensed AI training is lawful. It addressed the parties and evidentiary record before the court.
Practical record to retain: Acquisition records, the intended training use, and evidence relevant to rights and market effects.
Read the source
GEMA v. OpenAI
11 November 2025 decision · Munich Regional Court
The court largely upheld GEMA’s copyright claims concerning memorized song lyrics reproduced through the chatbot. It rejected the text-and-data-mining defense for the reproductions at issue.
A training-data rights assessment also needs to consider reproduction in model outputs; a manifest alone does not establish which source caused an output.
Practical record to retain: Source and rights records, output evaluation evidence, and the basis for any claimed mining exception.
Read the source
Getty Images v. Stability AI
4 November 2025 decision · United Kingdom High Court
The High Court rejected the secondary copyright claim based on model weights in the circumstances before it and found limited trademark infringement involving generated watermarks. Training-in-the-UK copyright claims had been withdrawn.
This was not blanket clearance of model training. Output, trademark, territorial, and training questions need to be distinguished.
Practical record to retain: Training-source records, relevant territorial facts, and output checks for protected content or marks.
Read the source
Thaler v. Perlmutter
2 March 2026 · United States Supreme Court
The Supreme Court denied review of Thaler’s challenge concerning a work submitted as created without human authorship. That left the lower-court result in place; denial of review is not a new merits opinion.
Wholly machine-generated material and human-authored contributions need separate assessment. AI assistance does not automatically remove protection from a person’s creative contributions.
Practical record to retain: A human-contribution record showing selection, arrangement, modifications, and other creative work relevant to the claimed rights.
Read the source · Read Copyright Office guidance