Adobe built its generative model on licensed and public images

A training-data claim built into the pitch

Adobe's current Firefly product page states that its models are “trained on licensed content, such as Adobe Stock, and public domain content where copyright has expired,” and adds explicitly that Adobe does not train its own Firefly models on a Creative Cloud subscriber's personal content. That distinction, licensed and public-domain sourcing rather than an unspecified web scrape, is the basis for Adobe's separate claim that Firefly's own models are “commercially safe,” and it is a claim about training method, not an independent audit of the training set.

Where the feature actually lives

The generative-fill capability that made this training claim visible to ordinary Photoshop users works, per Adobe's own Generative Fill page, by letting someone select part of an image and either describe a replacement in a text prompt or leave it blank for the model to fill based on the surrounding pixels, producing several variations on a new, non-destructive layer. The feature sits inside an editor a huge number of people already use for other work, which is precisely what made it a mainstream introduction to this kind of tool rather than a specialist product aimed only at people already seeking out generative image models.

Where the commercial-safety claim narrows

Adobe's own current page states a real limit worth pausing on: “content generated in the Photoshop app may not be used for commercial purposes” under the terms described there, a narrower claim than the general “commercially safe” language used for Firefly's own models elsewhere on Adobe's site. Adobe's separate Generative AI User Guidelines, a living legal document rather than a dated release, states that “outputs from generative AI features are for informational purposes only,” placing the burden of verifying that an output is fit for a specific commercial use back on the person publishing it. Reading the marketing language and the guidelines together, rather than the marketing language alone, is what surfaces that gap.

Questions to ask before you adopt it

• Does the specific feature you are using carry the same commercial-use terms as the model it is built on, or a narrower one?
• Has anyone at your organisation read the current generative AI usage guidelines, not just the product page?
• What happens if a generated asset later turns out to resemble a real, identifiable person or a third party's mark?

A training-data claim is not a warranty, and a feature's presence inside a trusted editor does not automatically extend that editor's other guarantees to what the feature produces. Adobe's own pages draw that line; the practical work is checking which side of it a specific use falls on.