Intellectual property and licence clearance

What it does

It establishes that the organisation has the right to use what goes into its AI systems and to exploit what comes out. Before a dataset is used for training, a model or component is adopted, or a significant generated work is released, someone confirms the rights position: who owns it, what the licence permits, whether an exception is being relied on, and whether the organisation can own or warrant the result. Legal sets the standard and decides doubtful cases; data, model and product owners carry out the checks. It operates at intake of data and components and before release of outputs.

How to implement it

• Record the source and licence of every dataset, model and component, and check the terms against the intended use.
• Review training and fine-tuning data for third-party works, and honour rights holders' opt-outs.
• Obtain warranties and indemnities from providers, and read their terms on outputs.
• Scan generated code for copied fragments and licence obligations.
• Check significant generated text and images for similarity to existing works before commercial use.
• Record human contribution to work the organisation intends to own.

Evidence it operates

Design: a clearance procedure and criteria for what counts as significant. Operation: clearance records for datasets and models in use; licence entries in the bill of materials; scan results for generated code; examples of an asset rejected or replaced; provider indemnities on file.

Where it fits

At the intake of data and components, and again before release. It is the primary control for copyright infringement in training data, infringing AI outputs, model and dataset licence breach and uncertain ownership of AI-generated work. It records its findings through dataset documentation and the AI bill of materials.

Limits

The law is unsettled, so clearance yields a judgement about risk and not a certificate. The organisation cannot inspect what a provider's model was trained on. Similarity checks catch close copies and miss works that are substantially similar in ways a tool cannot measure. Checking every output is impractical, which is why the control is aimed at what is significant.