A nonprofit wrote the rulebook behind AI licensing claims

The system

Fairly Trained is a nonprofit that certifies generative-AI companies whose training data meets published consent criteria. Its own about page states its mission directly: 'There is a divide emerging between two types of generative AI companies: those who get the consent of training data providers, and those who don't,' and that Fairly Trained 'exists to make it clear which companies take a more consent-based approach.' The organization is led by chief executive Ed Newton-Rex, and its advisers include Siri co-founder Tom Gruber and composer Max Richter; its listed supporters include Universal Music Group, SAG-AFTRA and The Authors Guild. Fairly Trained is the standards body behind any individual company's certification announcement, not a party to those companies' own products.

What the documents establish

Fairly Trained's published criteria for its first certification, called Licensed Model or 'L' certification, require that all training data used for a certified model fall into one of four categories: data provided under a contractual agreement with a rights-holding party; data available under an open license appropriate to the use case; data in the global public domain; or data fully owned by the model developer. The criteria state that licensing from an organization that itself licenses from creators, such as a record label or stock library, counts as consent for these purposes. Applicants must also show 'a robust process for conducting due diligence' into training-data rights and for keeping records of what data trained each model, and individual-model applicants must publicly disclose which of their models are and are not certified. Certification can cover an entire company, a single product, or one model, and is reevaluated annually after a submission and fee process the criteria document also describes.

Craft and rights

For a producer or label deciding which AI tools to use, Fairly Trained's certification functions as a licensing due-diligence shortcut rather than a guarantee: it tells a customer that Fairly Trained assessed the company's stated data sources against its published criteria, not that every training example was independently audited end to end. The certification is scoped to whatever was submitted, company, product or model, so a Fairly Trained mark on one product says nothing about a company's other, uncertified offerings. This entry does not describe any specific certified company's practices beyond what Fairly Trained's own criteria and process document state.

Outcomes and open questions

Because certification rests on a company's own submitted account of its data sources, reviewed against Fairly Trained's criteria rather than independently reconstructed from scratch, the durability of any certification depends on the annual reevaluation process holding up as a model's training data or licensing arrangements change over time.

• Does a certification cover the entire company, or only a specific named product or model?
• What category of the four permitted data sources does the certified developer's own disclosure rely on?
• Has the certification been reevaluated recently, or could underlying licensing arrangements have changed since?

A certification mark is only as informative as the criteria behind it, and Fairly Trained's own published criteria are the document worth reading before trusting the mark.