Licensed training data
How it works
The model provider enters into a contract with a rights holder (publisher, record label, stock library, platform) specifying which datasets may be used for training, for which uses, for how long and at what fee (one-off, subscription or revenue-share). The datasets are obtained via APIs, syndication channels or dedicated data dumps, and their origin is documented (data provenance) so that the right to use can be demonstrated. Opt-in/opt-out mechanisms let individual creators include or exclude their works. The data then enters a standard training pipeline (pretraining or fine-tuning), with settlement and usage reporting handled per the contract.
Problem solved
Training large models requires vast amounts of text, images, audio and video, much of which is protected by copyright. Scraping such content without consent creates legal (lawsuits, damages claims), reputational and regulatory risk, and leaves creators uncompensated. Licensed data provides a clear legal basis for use, predictable risk and a revenue stream for rights holders.
Components
Contract defining the scope of data use, permitted uses, duration and compensation.
Text, image, audio or video that the trainer has a confirmed right to use.
Documentation of the origin and legal status of data, enabling proof of a lawful source.
A tool letting creators voluntarily include or exclude their works from training.
A settlement model (one-off fee, subscription or revenue share) for rights holders.
Evolution
First high-profile content-for-training deals; in December 2023 The New York Times sued OpenAI and Microsoft, triggering a wave of licensing negotiations.
In June 2024 RIAA-coordinated labels sued Suno and Udio; in May 2024 OpenAI signed a licensing deal with News Corp.
From August 2025 the AI Act obligations for general-purpose models (training-data summary) apply; in November 2025 Suno settled with and licensed from Warner Music Group under an opt-in, artist-compensation model.