LLMs, GenAI, RAG, agents, evaluation · insight-1634
Read the provider's terms before training a model on its outputs
- Evidence type
- practitioner heuristic
- Difficulty
- easy
- Roles
- Applied AI, ML engineer, Data scientist
- Industries
- banking, finance
Summary: situation, action, reason, boundary
Situation
A plan to distil a hosted model's answers into an in-house small model is approved on technical grounds.
Action
Before generating training data, have the team and legal counsel read the use restrictions of the provider's commercial terms and of the licence of any open-weight base, record the decision with the terms version and date, and design a fallback data source.
Reason
Output use is a contractual matter, not a technical one. Anthropic's commercial terms (effective date shown June 17, 2025, as read 2026-10-12) restrict using the services to build competing products or to train competing AI models; other providers and licences differ and change.
Boundary
Any synthetic-data or distillation pipeline that uses a third-party model.
Symptom or risk
The finished model cannot be shipped or the contract is breached, after the labelling and training budget is spent.
Exceptions
Outputs from a model whose licence explicitly permits the intended use, which still deserves a documented check.
Trade-offs
The review adds days to the plan but costs far less than discarding a trained model.
Pitfalls
Assuming that the terms read last quarter still apply, or that an internal-only student is exempt without checking.
Lessons and concepts
Claims and their sources
Anthropic's Commercial Terms of Service (effective date shown June 17, 2025, section D.4) prohibit using the services to build a competing product or service, including to train competing AI models.