Model and dataset licence breach

What it is

Models and datasets come with terms. Many openly available models are licensed for research or non-commercial use only, or carry restrictions on particular uses, on redistribution, or on training other models from their outputs. Datasets have their own licences, which may not match that of a model trained on them. Hosted providers impose acceptable use policies. Developers accustomed to permissive open source software often assume the same applies and do not read the terms.

How it happens

• A model is downloaded from a public hub and put into a product without checking its licence.
• A dataset licensed for research is used to train a commercial model.
• A provider's outputs are used to train a competing model, contrary to its terms.
• A fine-tuned derivative is distributed under terms the base model forbids.

Why it matters

Breach exposes the organisation to claims and to loss of the right to use the component, which may sit at the centre of a product. The problem tends to surface at the worst moment, in due diligence for a sale or investment.

Warning signs

• No licence is recorded for the models and datasets in use.
• Research-licensed models appear in production systems.
• Nobody has read the provider's acceptable use policy.
• Derivative models are shared outside the organisation.