Licensing models for AI training data
A licence is a bundle of independent levers. Treat them separately; buyers will.
The levers
- Exclusivity
OptionsExclusive / non-exclusive / category-exclusive
Owner trade-offExclusivity limits future licences
- Term
OptionsFixed term / perpetual
Owner trade-offPerpetual rights are hard to unwind once models are trained
- Field of use
OptionsTraining only / evaluation only / any
Owner trade-offNarrow fields preserve other markets
- Redistribution
OptionsNone / to affiliates / sublicensable
Owner trade-offSublicensing loses visibility of end users
- Output terms
OptionsRestrictions on reproducing source content
Owner trade-offProtects the underlying product
Open licences are not a sale
Creative Commons terms such as NC (non-commercial) and SA (share-alike) were not written with model training in mind; their application to training is contested. Do not assume CC content you hold can be relicensed.
Have a dataset worth discussing?
Tell us what you own, which rights you can license and the commercial scope you have in mind.