The existing licenses cover AI training just fine, what we lack is sufficient legal precedent and enforcement. An AI product - more specifically, the model weights - is a derivative work of the original works used for training; AI training is a process of algorithmic compression of the originals.
Therefore, the resulting model should abide by all the license requirements imposed on the original - for example, if the model is trained on GPL code and can generate code, then any binary distribution should also be freely available for derivation in source format, and that includes all the algorithmically compressed training material (weights), which has become part of the model. If the source is AGPL, then that service cannot be made available on a website without disclosing said source and respective model weights.
Any other interpretation of the nature of copyright - which by definition, only covers human produced material - is just a variant of the proverbial "man that can't understand something because their paycheck depends upon them not understanding".
Therefore, the resulting model should abide by all the license requirements imposed on the original - for example, if the model is trained on GPL code and can generate code, then any binary distribution should also be freely available for derivation in source format, and that includes all the algorithmically compressed training material (weights), which has become part of the model. If the source is AGPL, then that service cannot be made available on a website without disclosing said source and respective model weights.
Any other interpretation of the nature of copyright - which by definition, only covers human produced material - is just a variant of the proverbial "man that can't understand something because their paycheck depends upon them not understanding".