A nonprofit began certifying AI models trained on licensed data

Fairly Trained's Licensed Model certification checks four criteria and has certified a dozen music-related companies since January 2024.

Historical event

January 17, 2024

First source published

January 17, 2024

Site publication

September 18, 2026

What happened

On 17 January 2024 the non-profit Fairly Trained launched its Licensed Model certification, awarding it at launch to nine generative AI companies spanning image, music and singing-voice generation, including Beatoven.AI, Boomy, Endel, LifeScore, Rightsify, Soundful and Tuney. Universal Music Group and the Association of American Publishers were named as supporting organisations at launch.

What the documents say

Fairly Trained's certification page, current as retrieved on 16 September 2026, sets out the Licensed Model requirements: all training data must be explicitly licensed by contract, available under an open licence, in the public domain, or owned by the developer, backed by a documented due-diligence process and record-keeping, with certification re-evaluated annually. The launch post states plainly that certification "will not be awarded to models that rely on a fair use copyright exception or similar," drawing a direct line between this certification and the training-data disputes described elsewhere in this file. The certified-models list, also current as retrieved, now includes music-specific entries such as Kits' voice and instrument models and Voice-Swap, alongside several of the original nine companies.

Why it matters for makers

A certification is only as strong as its verification process, and these documents describe what the certification requires without independently proving that any given holder's data sourcing was checked to the standard claimed. What the certification does provide is a specific, named commitment: a company holding it has represented to Fairly Trained that no copyrighted work was used without a licence, open licence, public-domain status or ownership, which is a narrower and more checkable claim than a general marketing statement about ethical AI.

What to check before you use it

A certification badge is a starting point for due diligence, not a substitute for it; check whether the specific product or model you intend to use is the certified unit, since certification can be awarded to an entire company, a single product, or one individual model within a larger catalogue. Confirm the certification is current, since it lapses without annual re-evaluation.

• Is the exact model or product I am using the certified unit, or a different model from the same company?
• Does the certification I am relying on remain current, given the annual re-evaluation requirement?
• What did the certified company itself disclose about its licensing sources beyond the fact of certification?

Certification narrows the question a maker has to ask from whether this was trained lawfully to whether this specific product met Fairly Trained's four criteria, which is useful but still worth checking rather than taking as a final answer.

Sources & reading trail

Fairly Trained's own announcement of the certification launch and the nine initially certified companies.

Source published: 17 January 2024 · Retrieved: 16 September 2026

Current criteria for the Licensed Model certification, including data-source and due-diligence requirements.

Source published: Not established · Retrieved: 16 September 2026

Current list of certified music-related companies and models as retrieved.

Source published: Not established · Retrieved: 16 September 2026

Papers, terms and official documents establish the record; the maker reading and the checks are Signal to Song editorial analysis. This retrospective draft does not imply the site published on the event date.

Continue reading

• Music groups published seven AI principles together
• Open models show two ways to disclose training data
• Suno assigns song ownership only on its paid tiers
• Browse the complete the archive

Published September 18, 2026, not on the date of the event described.