Adobe has so far marketed its artificially intelligent image generator Firefly as an ethical AI, trained primarily on its own library of licensed stock images (millions of photos, designs, and illustrations in the Adobe Stock library).

According to a new Bloomberg report, this clean image may not be as perfect as the Photoshop makers would like us to think. The training data for this "commercially safe" model may contain skeletons that are not as well known.

In the process of training Firefly's AI model, some of the images were provided by competitor Midjourney. The startup has not stated the source of the training data, but many suspect it came from images scavenged from the Internet without a license.

Adobe claims that only about 5% of the millions of images used to train Firefly fell into this category, all of which were part of the Adobe Stock library.

When Adobe first launched Firefly, it offered its corporate clients a disclaimer against claims of copyright infringement as a way to assure them that it was secure.

Adobe also marketed Firefly as a secure alternative to Midjourney and DALL-E.

Not all artists were so eager at the time, feeling coerced into agreeing to have their work used by the creative technology giant.

Despite the revelation that some images may have come from disreputable sources, Adobe stated that all non-human images were safe.

A spokesperson told Bloomberg: "All images submitted to Adobe Stock, including a small percentage of AI-generated images, do not contain IP, trademarks, recognizable characters or logos, or reference the artist's name undergoes a rigorous moderation process to ensure that they do."

The company appears to be taking a few more rigorous steps in its plans to build an AI video generator. Rumor has it that the company is paying artists per minute for video clips.

We have reached out to Adobe for comment on this story.