
The process wall
On 21 March 2023, Adobe introduced a generator built to sit inside existing production software rather than stand apart from it. The launch announcement states that Adobe's first model is trained on Adobe Stock images, openly licensed content and public domain content where copyright has expired, and is designed to generate content safe for commercial use. That training-data description is the process wall this record documents: a generator whose inputs are named, unlike models whose training sources are undisclosed, aimed at customers who cannot risk an unresolved rights question in delivered client work.
What the documents show
The announcement names three source categories for the initial model, Adobe Stock's licensed library, openly licensed content, and expired-copyright public-domain material, and states Adobe's intent to let customers later extend Firefly on their own creative collateral. Commercially safe is Adobe's own description of the design goal, restated on the current product page, retrieved 16 September 2026, as: Adobe Firefly models are commercially safe, a vendor claim tied to the stated training sources rather than an independent certification. The Firefly FAQ, also a living document as retrieved on 16 September 2026, adds a narrower and more checkable claim: Adobe does not train on Creative Cloud subscribers' personal content, while Adobe Stock contributor content is used in accordance with Stock Contributor licence agreements. The announcement also introduces a Do Not Train Content Credentials tag that remains associated with content wherever it is used, published or stored.
Authorship and limitation
The mechanism carrying Adobe's transparency claim is metadata, not enforcement: a Content Credential travels with a file as attached data, and the FAQ links it to the Coalition for Content Provenance and Authenticity, or C2PA, standard for recording an asset's origin and edit history. That is a credential, not a licence: it can indicate that AI was used or that a creator opted out of training, but stripping or ignoring the metadata is a separate technical possibility the documents do not address. Commercially safe describes the training-source category Adobe selected, not a guarantee that no output could ever resemble an existing work.
What survives
A maker relying on Firefly for client work can check the FAQ's training-source claim, whether Content Credentials survived a given export, and whether Adobe's promised Stock-contributor compensation model, a future commitment in the announcement, has since been detailed. Commercially safe should be quoted as Adobe's claim, not restated as settled fact.
- Check whether an exported Firefly image still carries its Content Credential after passing through other software.
- Read the FAQ's exact wording on Creative Cloud personal content versus Adobe Stock contributor content before generalising about training.
- Ask which of the three named training-source categories a specific Firefly model version actually used, since the FAQ describes this as evolving beyond the initial model.
The announcement's real significance for a commercial maker is not the image quality it claims but that it names its sources at all, turning a training-data question that is usually unanswerable into one with a specific, citable, checkable statement.
Sources & reading trail
States the training-source categories, the commercially safe design goal, and the Do Not Train Content Credentials tag.
Source published: 21 March 2023 · Retrieved: 16 September 2026
Living FAQ distinguishing personal Creative Cloud content from Stock contributor content in training, as retrieved 16 September 2026.
Source published: Not established · Retrieved: 16 September 2026
Current vendor page restating the commercially safe claim and describing Content Credentials, retrieved 16 September 2026.
Source published: Not established · Retrieved: 16 September 2026
Papers, documentation, rulings and the makers' own records establish the note; the authorship reading and what-survives are Creative AI editorial analysis. This retrospective draft does not imply the site published on the event date.