Copyright and Training Data
Whether training a model on work you didn't license is lawful — unresolved, consequential, and the industry is shipping into the uncertainty at scale.
When not to use it
- (It's an unresolved question, not a technique.)*
- As a settled matter, in either direction. Courts have split. Both arguments are serious.
- Assuming your provider's indemnity covers output similarity. Training and output are separate questions.
- Assuming robots.txt settles anything. It's a convention, not a law, and copyright is separate from access.
- Assuming code is like text. Open source licences are explicit and copyleft raises a distinct question.
Reach for something else instead
- Licensed-data models — provenance as a product, at some quality cost.
- Public domain and permissively licensed corpora — clean, smaller, weaker.
- Provider indemnification — moves the risk contractually; doesn't answer the question.
- Licensing deals — the commercial answer arriving before the legal one.
Sources & further reading
- Carlini et al. (2023), Extracting Training Data from Diffusion Models — memorisation is real and rare, which matters for the "no copies stored" argument.
- Henderson et al. (2023), Foundation Models and Fair Use — the careful legal analysis; fair use is not the blanket defence it's assumed to be.
- Lee, Cooper & Grimmelmann (2024), Talkin' 'Bout AI Generation: Copyright and the Generative-AI Supply Chain — where liability attaches at each stage. The clearest map available.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Collapsing training and output into one question. They're distinct and can resolve differently.
- Citing "humans learn from books too." It's a good intuition and not a legal argument — scale and market effect are what the doctrine actually weighs.
- Claiming models don't store copies. Memorisation is documented; rare isn't zero.
- Missing that a licensing requirement would entrench the biggest labs and eliminate open competitors.
At a glance
FieldSafety & Ethics
Statusunresolved; courts have split
The four-factor question that matterstransformativeness, and market effect
Training vs. outputtwo separate questions, routinely collapsed
The awkward interactionprotecting creators may entrench the largest labs
DifficultyBeginner
Flashcards for this concept
Question
—
Answer
—
1 / 4
Often compared with
Training vs. output infringement — one asks whether reading the corpus was lawful, the other whether what came out is too close to something specific. A model can win the first and lose the second.