AI models are typically trained on large data sets, often mined from copyrighted works. Without explicit permission from the copyright holder, e.g. in the form of licences, or if the training is not supported by a provision in copyright law, this constitutes an infringement of copyright.

It has not yet been established whether the existing provision for text and data mining is applicable to generative AI models. However, it is undisputed that training carried out by commercial entities within the EU in cases where copyright holders disallow use and no valid licence exists, is not permitted. Several Swedish licensing and collecting societies representing affected copyright holders have made collective reservations – opt outs – on account of their members.

Against this background, a copyright assessment should always be made before using generative AI.

Users should consider the following points:

• Identify what kind of training data the supplier of the AI model has used – and will be using – when developing their service and whether any permits and licences from copyright holders exist. If there is a risk of unlawful training, the service should not be used without permission. Permission can normally be obtained from collecting societies that represent copyright holders.
• Identify in what way the supplier of the AI model will be using the data that is fed into the system via a prompt, or whether the supplier will be claiming any rights to the created content. This should be included in the terms of use. However, terms are often vaguely formulated. It is therefore important to make your own assessment.
• Identify whether the supplier has been involved in a copyright lawsuit or if they are waiving liability, which may be an indication against using the service.

If your risk assessment suggests that the service should not be used, it is recommended to instead select a licensed model, or models that are proven to have been trained on copyrighted works with permission.