DataPacts

Training Data Certifications: Fairly Trained, Credtent and Other Labels

Checked

At a glance

A training data certification is a private label, issued by an organization and not by a regulator, that says an AI company or a content catalog meets the issuer's own conditions about how training data was obtained or where content came from.

Independent reference. Every buyer is covered on the same terms, every figure carries its source and check date, and no buyer pays to be listed or placed higher.

Request an independent introduction

Tell us what data your company holds. We reply with the buyers that fit and what their terms leave out. Your details reach a buyer only after you agree to a named introduction.

IndependentYour details stay with us until you agreeNot selling is a valid outcome

Only one issuer found for this page, Fairly Trained, certifies a developer's training data sourcing. Its Licensed Model certification, launched in January 2024, asks that all training data be licensed, openly licensed, in the public domain or owned by the developer.

This page reads three issuers' own pages, Fairly Trained, Credtent and C2PA, with the criteria and fees they publish as read on 2026-11-17, and states what such a label leaves unproven.

No label is recommended. The labels matter to a data holder because they describe what a developer says it checks before it accepts data in AI training data licensing.

What a training data certification is and who issues it

A certification is a claim by a named issuer that an applicant met published conditions at a date, and it binds no one else. It is not a finding by a court or an agency, and it does not change what the law requires of a developer or a licensor.

Three kinds of label appear on this page:

  • A certification of an AI company's training data practices (Fairly Trained).
  • A badge or tag on a work or a catalog (Credtent, C2PA).
  • A standard or audit that certifies nothing (the Data & Trust Alliance standards and the Data Provenance Initiative).

The legal duties sit elsewhere and are public, such as the EU training content summary and state disclosure laws. DataPacts certifies nothing and recommends no label. The private record of who held which right, which any label rests on, is the subject of training data provenance.

Four labels and standards compared by what they check

Only one of the four items below is a certification of training data practices, and the others label content or describe a dataset. The table compares them as of 2026-11-17, in alphabetical order of label.

Label or standard Issuer What the issuer's page says it covers Certifies training data practices? Fee stated on the page read Source
Content Credentials (C2PA) Coalition for Content Provenance and Authenticity origin and edits of digital content; not DRM; no attribution in the core specification no; training assertion removed in version 2.0, January 2024 not applicable (open standard) c2pa.org
Creative Origin badges and certification Credtent how a work was created: human, AI-assisted, AI-composed, legacy, multi-origin no; describes a work's origin, not training data badges described as free; certification fee not stated credtent.org; registry.credtent.org
Data Provenance Standards v1.0.0 Data & Trust Alliance 22 metadata fields: Source 6, Provenance 8, Use 8; published 2024-07-09 no; voluntary standard not applicable dtaalliance.org
Licensed Model (L) certification Fairly Trained training data sources, due diligence and record keeping of the applicant yes, per its own page yes, by revenue band fairlytrained.org

Fairly Trained is the single issuer this page found that certifies a developer's training data sourcing, so "other labels" in the title means the adjacent labels in the table. No other training data certification is recorded in this reference, and the page does not claim that none exists.

Fairly Trained's Licensed Model certification: criteria, process and fees

Fairly Trained, launched in January 2024, certifies under one program its own page describes, the Licensed Model (L) certification, which asks that all training data be licensed, openly licensed, in the public domain or owned by the developer. Its page, read on 2026-11-17, sets five criteria:

  • Scope: the applicant may certify a whole company, a single product or service, or an individual model.
  • Data sources: training data is provided for training under a contract with a party that has the rights, available under an open license suited to the use, in the public domain globally, or fully owned by the developer. A license from an organization that itself licenses from creators counts, and third-party, open or synthetic data meet the same requirements.
  • Due diligence: a robust process for checking training data, including checks on the rights position of each data provider.
  • Record keeping: a robust process for keeping records of the data used in each training.
  • Disclosure: individual-model applicants state publicly which of their models are and are not certified, in wording agreed with Fairly Trained.

The page describes four steps, the last of which repeats each year:

  1. An online application, after which Fairly Trained guides the written submission.
  2. The written submission, sent with the submission fee, which Fairly Trained may follow with requests for more information.
  3. If approved, payment of the annual certification fee.
  4. A reevaluation of the certification every year.

The table gives the fee schedule by annual revenue as Fairly Trained's page stated it on 2026-11-17, in USD.

Annual revenue Company or product: submission fee Company or product: annual fee Individual model: submission fee Individual model: annual fee
Under USD 100,000 USD 150 USD 500 USD 100 USD 250
Under USD 500,000 USD 150 USD 1,000 USD 100 USD 500
Under USD 5 million USD 250 USD 2,000 USD 150 USD 1,000
Under USD 10 million USD 350 USD 4,000 USD 200 USD 2,000
Over USD 10 million USD 500 USD 6,000 USD 250 USD 3,000

Source: fairlytrained.org/certifications, read 2026-11-17. The page carries no date, so the schedule is the issuer's published one and not a confirmed current price. The page read states no certified count and no exclusions, and this page repeats neither a launch count nor a legal status.

Credtent's badges and certification: what its own pages say

Credtent states on its own site that it tags content for provenance and offers Creative Origin certification, and the Creative Origin badges it publishes describe how a work was created, not which data trained a model.

The registry page, read on 2026-11-17, lists five types: Human-Composed, AI-Assisted, AI-Composed, Legacy (a work made before November 2022 without generative AI tools) and Multi-Origin.

Its headline offers the badges free, and the registry page names no verifier and no verification method for the badges. Credtent describes itself as a neutral, third-party utility, which is a self-description.

Credtent also offers rights holders a free opt-out notice, with follow-ups at 30, 60 and 90 days, and offers AI companies a pre-cleared chain of title and disclosure pack, whose license covers only the rights holders who warranted them and the named works, by Credtent's statement.

It states that it is a Delaware Public Benefit Corporation. Its licensing terms belong on the terms comparison, and the notice mechanics on the page about opting out of AI training.

C2PA and industry metadata standards: labels for content, not training sets

C2PA's Content Credentials record the origin and edits of a piece of digital content, and they are not a certification of the data an AI model was trained on.

C2PA's FAQ, read on 2026-11-17, states that Content Credentials are not DRM, that the core specification does not support attribution of content to individuals or organizations, and that they record provenance without limiting use. Version 2.0 of the specification, from January 2024, removed its Training and Data Mining assertion.

The Data & Trust Alliance published Data Provenance Standards v1.0.0 on 2024-07-09, with 22 fields in three standards (Source 6, Provenance 8, Use 8) that describe a dataset's origin and permitted use. It is a voluntary standard with no certification step.

The Data Provenance Initiative is a different group, a volunteer research collective that audits dataset licenses and consent signals and certifies nothing. How these standards compare with datasheets and data cards is covered under dataset documentation.

The asset-level permission record that replaced the removed C2PA assertion is covered on the parent page.

What a certification proves and what it leaves open

A label proves what its issuer checked on the date it was issued, and the issuers' own pages say little about what they checked beyond the applicant's own process. The pages read say the following:

  • Fairly Trained's criteria concern the applicant's data sources and its due diligence and record keeping processes.
  • C2PA records the history of a content item and not limits on its use.

The pages read do not state the following:

  • Whether Fairly Trained inspects datasets beyond the written submission.
  • The certified count, the list of certified models and any exclusions.
  • Who verifies a Credtent badge, by what method and at what fee.
  • The date through which any named model holds certified status.

A label is not a regulator's finding and replaces no legal duty. Article 53(1)(d) of the EU AI Act (Regulation (EU) 2024/1689) requires providers of general-purpose AI models to publish a summary of training content, applying from 2025-08-02, whether or not any label is held.

No court record is cited here. The disclosure laws themselves are laid out under training data transparency rules.

What a label changes for a company licensing data

A label sits on the developer's side of the deal, and for a seller it matters because Fairly Trained's criteria ask the applicant to check the data provider's rights. Three consequences follow for a seller:

  • A buyer seeking the L certification may ask for evidence of the seller's rights, which is the chain of title.
  • The contract can state what may be said publicly about the deal.
  • A label on a buyer does not change the seller's own rights audit or what the seller warrants.

The rights record that a due diligence check reads is built step by step on the page about the chain of rights for training data. The seller process pages cover the diligence a buyer runs and the contract that records it.

Benefits, risks, downsides and alternatives of seeking or relying on a label, including not licensing

A label can give a buyer or a seller a shared, public reference point, and it can also be read as more than the issuer says it covers.

Benefits:

  • A published, dated set of conditions, as in the L criteria.
  • A public statement, for individual-model applicants, of which models are certified.
  • A common vocabulary in a negotiation about sources and records.

Risks:

  • Reading a label as legal clearance.
  • Reading a Credtent origin badge as a training data audit.
  • Relying on outdated status, since Fairly Trained reevaluates each year.

Downsides:

  • Submission and annual fees for the applicant, in the schedule above.
  • Verification methods left unstated on some pages read.
  • No regulator or court recognition recorded here.

Alternatives:

  • Rely on a rights audit and contract warranties.
  • Publish a dataset card and a provenance record.
  • Rely on the legal public summary where a law applies.
  • Not license at all, which is a valid outcome.

Where certification meets the chain of rights, the law and the seller process

Certification is one private layer above the seller's own records, the public disclosure duties and the contract, and the pages below take each part.

Frequently asked questions

What is the Fairly Trained certification?

Fairly Trained is a certifier of AI companies' training data practices, launched in January 2024. Its Licensed Model (L) certification is described on its own page, with a fee schedule by annual revenue, as read on 2026-11-17.

Does a certification mean a model uses no copyrighted data?

No, the L criteria concern licensed, openly licensed, public domain and owned data, and the page read states no audit of the data itself. A licensed work can still be copyrighted, and the label is the issuer's claim about the applicant's process.

Is C2PA a training data certification?

No, C2PA records the origin and edits of content. Its specification removed the Training and Data Mining assertion in version 2.0, January 2024.

Does DataPacts certify anything?

No, DataPacts certifies nothing and recommends no label. It reads each issuer's own pages and dates what they state.

How this page is sourced and dated

The issuers' own pages were read on 2026-11-17, together with the founder's launch post of 2024-01-22 and the Data & Trust Alliance standard page. This page is information, not legal advice.