On Friday, Bloomberg reported that Adobe’s AI image generator Firefly included AI images from competitors in its training data — a case of AI learning from AI.
The report brought a new dimension to Adobe’s claims of Firefly’s ethicality. In public-facing articles, Adobe differentiated Firefly from competitors like DALL·E 3, Stable Diffusion, and Midjourney by emphasizing Firefly’s “commercially safe” training data.
“Adobe has established the AI ethics principles of accountability, responsibility, and transparency,” Adobe wrote in one post.
Firefly drew from licensed Adobe Stock images, the company explained, plus images in the public domain. Adobe even created a bonus compensation plan for artists whose work went into the first release of Firefly.
Some of those artists submitted images generated by Midjourney and other rival AI, and were compensated by Adobe for their input, according to Bloomberg’s report.
Symbol of an ouroboros, or serpent eating its own tail, generated by Adobe Firefly in response to Entrepreneur’s prompt: “A snake is eating its own tail in an infinity symbol against a background with mountains, trees and cloudy sky.” Credit: Adobe Firefly
Though the artists had to note that their work was created with AI, they did not have to tell Adobe which generator they used.
This means that, even if Firefly isn’t actively scraping the Internet without permission, AI image generators like Midjourney might be. And Firefly could be trained on those Midjourney images.
Related: JPMorgan CEO Jamie Dimon Says AI Could Impact ‘Every Job’
Multiple Adobe employees told Bloomberg about an internal clash over the ethics of training Adobe’s AI with AI-generated images from rivals. Even in the early stages of developing Firefly, some employees disagreed with the company’s decision to include AI images in Firefly’s training data.
In Adobe Stock, the database used to train Firefly, 57 million are labeled AI-generated, per Bloomberg. That’s 14% of the total images in the database.
Adobe responded to the claims by stating that only 5% of the images used to train Firefly came from AI images created by other platforms.
“Every image submitted to Adobe Stock, including a very small subset of images generated with AI, goes through a rigorous moderation process to ensure it does not include IP, trademarks, recognizable characters or logos, or reference artists’ names,” a company spokesperson told Bloomberg.
Related: OpenAI Reportedly Used More Than a Million Hours of YouTube Videos to Train Its Latest AI Model
The race to develop the next big AI has intensified the pressure companies face to find new sources of training data. According to reporting from The New York Times, OpenAI may have trained its latest text-to-video AI generator Sora on YouTube videos, and Google may have been doing the same thing.
Adobe’s products have captured the bulk of global market share for leading graphics software, with Adobe Photoshop, InDesign, and Illustrator comprising more than 80% of the market, according to Statista. Canva has 10.26% of the global market, per the same source.
A Bloomberg report from earlier this week shows that Adobe has started paying its network $2.62 to $7.25 per minute of recorded video for fresh data to use in its own Sora competitor.
Related: Authors Are Suing OpenAI Because ChatGPT Is Too ‘Accurate’