Adobe AI: What Firefly Is Trained On and What Terms Say

Adobe Sensei and Firefly explained: stated training data, terms on customer files, partner models, and what Content Credentials can show.

Guide: AI Tools and IndustryLast verified: October 20265 min read

Adobe's generative AI is centered on Firefly, a model family announced as a beta on March 21, 2023, and on Adobe Sensei, a machine learning framework announced in 2016. Adobe states that Firefly's own models are trained on licensed, openly licensed, and public domain content, and that its terms bar training generative AI on customer content. Since April 2025 the Firefly app has also offered models from other companies, which is a separate matter from Adobe's own models.

Key takeaways

  • Adobe Sensei, announced in November 2016, predates the generative tools and covers machine learning features across Adobe products.
  • Firefly launched as a beta on March 21, 2023, with a first model trained on Adobe Stock, openly licensed, and public domain content.
  • Adobe's terms published October 3, 2025 say it will not use local or cloud content to train generative AI models, with an exception for content submitted to Adobe Stock.
  • Partner models from OpenAI, Google, and Black Forest Labs joined the Firefly app from April 24, 2025.
  • Content Credentials record how an item was made, but they help only while the label stays attached.

Two generations of Adobe AI

Adobe's machine learning work was branded Adobe Sensei when it was announced at Adobe MAX in early November 2016, described at the time as a framework and set of intelligent services behind Creative Cloud and Document Cloud. Features such as background removal and masking in editing software belong to this earlier layer. Adobe states that product improvement based on usage data and content characteristics uses "machine learning (NOT generative AI)," and that users can opt out of desktop product improvement programs.

The generative layer arrived with Firefly. Adobe's March 21, 2023 announcement introduced a beta focused on images and text effects, with a first model trained on Adobe Stock images, openly licensed content, and public domain content whose copyright has expired. The Firefly app, presented as an all-in-one home for Adobe's models, was announced on April 24, 2025, together with a generally available Firefly Video Model.

What the terms say about customer files

The question most users ask is whether their own files train the models. Adobe's Terms of Use, published and effective October 3, 2025, state in section 2.2(F) that Adobe will not use local or cloud content to train generative AI models, and repeat the point in section 4.3. Both carve out content submitted to the Adobe Stock marketplace, which is governed by its own contributor terms. Section 4.3 also says the license Adobe holds to operate its services does not give permission to train generative AI models on customer content.

The wording has a history. In June 2024 a re-acceptance of the terms prompted a backlash over language about accessing content and using machine learning. Adobe's blog post of June 10, 2024 said it had never trained generative AI on customer content, did not claim ownership of customer work, and planned clearer language by June 18, 2024. Adobe also said it scans uploaded content for child sexual abuse material, with human review of flagged items, and does not scan content stored locally on a computer.

Own models and partner models

Adobe's April 24, 2025 announcement described Firefly as including all of its "commercially safe" creative models, and added partner models: OpenAI's image generation, Google's Imagen 3 and Veo 2, and Flux 1.1 Pro, with Ideogram, Luma, Pika, Runway, and fal models announced to follow. Adobe's claim of safety refers to its own models. Partner models are trained by other companies on data Adobe does not describe, so the training-data statement for Firefly does not carry over to them. Adobe said enterprise customers could choose whether to enable partner models across their organization.

"Commercially safe" is Adobe's term and not a legal category. Whether a given output is protected by an intellectual property indemnity depends on the customer's contract and plan. Coverage for Adobe's own models should not be assumed to extend to every partner model, and Adobe's terms and help pages change, so the current terms for the specific plan and model are the thing to read.

Content Credentials: what a label can and cannot show

Adobe states that Content Credentials are automatically attached to AI-generated content and indicate whether a Firefly model or a partner model created it. The format is an open standard: the Coalition for Content Provenance and Authenticity (C2PA) publishes the technical specifications for certifying the source and history of media. The Content Authenticity Initiative, founded by Adobe in 2019 and still led by it, builds open-source tools around the standard.

Content Credentials are cryptographically signed, which makes changes to the content or the recorded information detectable. They can include who produced an item, when, and with which tools, and identity is optional. The limit is that a credential is metadata. A screenshot or a platform that strips metadata can separate an image from its record, so the absence of a label does not show that an image is real or unedited.

Checking an Adobe AI claim

This short list is an illustrative way to test statements about Adobe AI.

  • Check the date: the Terms of Use, product descriptions, and partner model lists change.
  • Check the model: Firefly's own model or a partner model.
  • Check the plan: personal, team, or enterprise terms can differ on indemnity.
  • Check the data: customer content, Adobe Stock content, and public data are treated differently.
  • Check what the label proves: a Content Credential shows a recorded history, not truth.

FAQ

Does Adobe train generative AI on files stored in Creative Cloud?

Adobe's Terms of Use effective October 3, 2025 say it will not use local or cloud content to train generative AI models, except for content submitted to Adobe Stock. Adobe's 2024 blog post said the same thing about past practice.

What was Firefly originally trained on?

Adobe's March 2023 announcement said the first model was trained on Adobe Stock images, openly licensed content, and public domain content where copyright has expired.

Are partner models in Firefly made by Adobe?

No. Models from OpenAI, Google, Black Forest Labs, and others are built by those companies and offered inside the Firefly app. Adobe's commercially safe description applies to its own models.

Can Content Credentials prove an image is authentic?

They can show a signed record of how a file was made and edited. They cannot show anything once the data has been removed, so they are one signal and not proof.

Last verified: October 2026

Sources

  • Adobe Unveils Firefly, a Family of New Creative Generative AI, Adobe news release, March 21, 2023
  • Updating Adobe's Terms of Use, Adobe Blog, June 10, 2024
  • Adobe General Terms of Use, Adobe, published and effective October 3, 2025
  • Adobe Revolutionizes AI-Assisted Creativity with Firefly, Adobe news release, April 24, 2025
  • Adobe Sensei announcement, Adobe MAX press release, November 2016
  • How Content Credentials work, Content Authenticity Initiative
  • C2PA Specifications, Coalition for Content Provenance and Authenticity

New to artificial intelligence?

Start with the basics, then follow the guides to tools, rules, and effects on society.

Start with AI Basics