Fair Use for Lawfully Sourced Training Data
Tags: Regulatory
TL;DR
- A U.S. federal court found that Anthropic’s training on lawfully obtained copyrighted works was fair use.
- Use of pirated materials was not protected, and the decision addressed training data only—not outputs.
Why it matters for HK marketers: This signals how courts may separate lawful training from infringing outputs, shaping vendor risk allocation and your indemnity posture.
What the court held
- Fair use for lawful sources: Training on lawfully obtained works deemed transformative.
- No cover for piracy: Use of pirated content was not protected.
- Scope limited to training: The ruling did not adjudicate whether outputs infringe.
Practical implications
- Vendor diligence: Demand provenance proof for training datasets.
- Contract design: Separate warranties and indemnities for training vs. outputs.
- Policy stance: Reinforces the need for lawful sourcing and auditability.
The court labeled lawful training use 'spectacularly transformative.'
The fair-use finding applied to training materials, not outputs.
The case settled and the fair-use decision was not appealed.
So what for marketers
Insist on lawful sourcing warranties and audit rights. Structure indemnities to cover third-party claims tied to both training data and outputs.
← Back to Knowledge Base