Adobe is trying to prove a point at the center of the generative AI debate: powerful AI tools do not have to be built on disputed copyrighted material. Its Firefly image model, released one year ago and integrated into Photoshop, was designed around licensed data and creator concerns rather than broad scraping of the open web.
The company’s leaders argue that this is not only an ethical position. They say it is also becoming a practical business strategy, especially for customers that need legal certainty, professional safety, and clearer signals about how AI-made content was created.
A different answer to the AI training fight
Since the generative AI boom began, one of the central disputes has been how large AI models get their training data. Some technology companies, including OpenAI, have claimed it is “impossible” to train AI without collecting copyrighted data from the internet. Artists and writers have argued the opposite: that their work has been taken without permission or compensation.
Adobe has taken a more unusual position for a major AI company. It says Firefly was trained on content with an explicit license allowing AI training. According to Ely Greenfield, Adobe’s chief technology officer for digital media, much of that training data comes from Adobe’s own stock photo library.
The company also says it offers creators extra compensation when their material is used to train AI models. That makes Firefly part of a broader push toward a license-based model, where creators are paid when their work is included in training datasets.
David Wadhwani, president of Adobe’s digital media business, frames the issue as a question of restraint as well as capability. “We worry that the industry, Silicon Valley in particular, does not pause to ask the ‘how’ or the ‘why.’ Just because you can build something doesn’t mean you should build it without consideration of the impact that you’re creating,” he says.
Why licensed training data matters
Scraping online data has become one of the most controversial practices in AI. OpenAI, Stability.AI, Meta, and Google are facing numerous lawsuits over AI training data. Technology companies argue that publicly available data can be used, while creators are pushing for licensing and compensation.
Adobe’s approach is different because it narrows what Firefly is allowed to learn from. Human content moderators review the training data to remove objectionable or harmful content, known intellectual property, and images of known people. The company says it has licenses for everything its products train on.
That choice came with risk. More data has traditionally helped developers build stronger models, and scraping the internet is a cheap way to collect large training sets. Greenfield says Adobe did not know at the start whether Firefly could meet customer needs without using scraped web data.
“To be honest, when we started with Firefly with our image model, we didn’t know whether or not we would be able to satisfy customer needs without scraping the web,” says Greenfield.
“And we found we could, which was great.”
For Adobe, the bet is that professional users value trust, control, and rights clarity. In Photoshop, Firefly lets users fill areas of an image through text commands. That keeps generative AI close to an existing creative workflow instead of making it a separate machine for producing finished images with little user direction.
Moderation starts with what the model has seen
Firefly’s training strategy also shapes its content moderation. Generative AI systems are difficult to control, and developers do not always know why models produce particular images or text. Models trained on broad internet datasets can reflect what those datasets contain, including copyrighted content, personal data, and harmful material.
Greenfield says Adobe’s model has not seen pictures of Joe Biden or Donald Trump, and that it has no news content or famous people in its training data. It also has not been trained on copyrighted material such as images of Mickey Mouse.
“It just doesn’t understand what that concept is,” says Greenfield.
Adobe also applies automated content moderation when Firefly creates images. The model is prohibited from creating news stories or violent images, and some artists’ names are blocked. Greenfield says it currently takes around 10 seconds for Adobe’s content moderation algorithms to check model outputs.
Firefly-generated content also comes with labels that indicate it was created using AI and show the image’s edit history. Adobe has been a vocal advocate for this kind of labeling, especially because knowing who made a piece of content, and how it was made, is important during a critical election year.
Business demand is part of the story
Adobe’s argument is not only about avoiding harm. The company says responsible AI can still work commercially. Wadhwani says Firefly-powered features are among Adobe’s most popular, and 90% of Firefly’s web app users are entirely new customers to Adobe.
The company is also working with businesses that want to use generative AI with their own intellectual property. Adobe has teamed up with companies such as IBM, Mattel, NVIDIA and NASCAR, allowing those companies to use the tool with their intellectual property. It is also working on audio, lip synching tools and 3D generation.
At the same time, Adobe’s strategy points to a larger challenge for the AI industry. If companies want models that businesses and creators can trust, the origin of training data, the way outputs are moderated, and the labeling of AI-generated content all matter.
Adobe started the Content Authenticity Initiative with the New York Times and Twitter, now X. The initiative now has over 2,500 members. Adobe is also part of developing C2PA, an industry standard label that shows where a piece of content came from and how it was created.
Claire Leibowicz, head of AI and media integrity at the nonprofit Partnership on AI, says Adobe’s approach supports social goals such as fighting misinformation while also serving business goals such as preserving creator autonomy and attribution.
“The business mission of Adobe is not to prevent misinformation, per se,” she says. “It’s to empower creators. And isn’t this a really elegant confluence of mission and tactics, to be able to kill two birds with one stone?”
For Adobe, Firefly is now evidence that a generative AI product can be built around licensed data, compensation, moderation, and labels without abandoning growth. Wadhwani puts the business case directly: “I think our approach has definitely been good for business,” he says.