Licensing AI Models and Data
The contract is where IP risk gets managed. Traditional IP indemnities break down for AI because they exclude modifications, combinations and out-of-scope use - all inherent to AI. Licensees should insist on minimum performance metrics, warranties and a trial period; the Designova case shows the cost of undefined ownership.
Who owns the data, who may use it, and who pays when it goes wrong. The contract is where IP risk gets managed.
What licensing terms regulate:
- Designating model components as trade secrets.
- Protecting components → limiting use rights, designating confidential information, restricting its use.
- Assignment rights in model evolutions from one party or the other.
- Licence and use rights per model component; rights in the T&Cs.
- Liability and indemnification
Traditional IP indemnities exclude → modifications, unauthorised combinations, use beyond authorised scope. For AI models these exceptions break down, because by design the model must be trained (a modification), must be combined with training and production data, and may evolve beyond a pre-set scope over time.
What licensees should insist on: many models are still experimental → demand minimum performance metrics (accuracy, reliability, robustness) · contractual warranties and indemnities against underperformance · testing and validation before licensing and on an ongoing basis · a trial period helps.
GenAI outputs: if an AI generates a work → who owns the copyright, the developer or the user? Scraping public sources risks IP misuse, and AI makes infringement harder to detect and enforce when works are similar but not identical to existing ones.
Ownership of AI-generated product designs was ambiguous → the fix was a governance framework with policies explicitly defining ownership attribution and vendor agreements specifying IP rights, plus regular risk assessments. Failing to define ownership risks → misaligned views, conflicting ownership claims, protracted disputes draining resources.