A certification mark targets AI training data, not book bylines

The writer's problem

A reader trying to tell which AI companies licensed their training data from those relying on a fair-use defense has had no easy public marker to check, since the underlying training runs are rarely disclosed. On 17 January 2024, a new nonprofit, Fairly Trained, launched a certification aimed at exactly that gap.

What the documents show

Fairly Trained's own launch post states it began with nine already-certified companies spanning image, music and singing-voice generation, and that its Licensed Model certification will not be awarded to a model that relies on a fair-use exception rather than an actual license for its training data. The organization's current certification criteria require that all training data fall under a contractual license, an open license, the public domain, or the developer's own ownership, plus a documented due-diligence process and ongoing record-keeping, reevaluated annually for a fee that scales with the applicant's revenue. Its about page lists supporters including the Authors Guild and the Association of American Publishers, and advisers including Maria Pallante, president of the Association of American Publishers. This is the organization's own account of its criteria and supporters, not an independent audit of any certified company's actual training practice.

The editorial choice

Fairly Trained built its first certification around consent for training data specifically, not compensation, credit, or a book's own authorship; its published FAQ states this focus was a deliberate starting point, with the organization saying it plans to address compensation and credit questions in future certifications rather than in this one. A writer citing a certification should describe it as evidence of a licensing claim the company chose to have checked, not as a court's finding that any company's training was lawful.

What stays with the author

The certification says nothing about whether a specific book, once written, used AI at any stage, and nothing about an individual author's own disclosure obligations to a publisher or reader; those remain separate questions the certification does not touch. Fairly Trained's own FAQ states that certification can be rescinded if a company's practices change, meaning a certification observed on one date is not a permanent guarantee.

• Is this a full-company, product, or single-model certification, and does the AI tool in question fall inside it?
• Does the certification claim consent for training data, or something broader about compensation or credit?
• How recently was this certification reevaluated, given that Fairly Trained states it can be withdrawn?

Fairly Trained's mark addresses the training-data side of the AI-and-authorship debate, leaving the separate question of how any resulting book is written and credited entirely where it already sat, with the author.