According to a Bloomberg report, Adobe has been training its Firefly image generator with images from rival generator Midjourney and presumably other image AIs. This is awkward because Adobe advertises that Firefly is legally and ethically sound.
Adobe used AI-generated images from competitor Midjourney to train its Firefly image AI, Bloomberg reports, citing statements on Adobe's Discord discussion group. This contradicts Adobe's claims that it only trains Firefly with content it owns the rights to or that is in the public domain.
Adobe had previously stressed that this approach differentiated it from competitors such as Midjourney. But according to the statements, AI-generated images, including those from Midjourney, make up about five percent of Firefly's training data.
AI images enter Firefly training via Adobe's stock database
The AI images entered Firefly's training via Adobe's own stock database, where they have been uploaded and sold for some time. Last November, Adobe faced criticism for not moderating AI images related to the Israel-Hamas war on its platform.
An Adobe spokesperson told Bloomberg that all images uploaded to Adobe Stock, including AI-generated images, go through a rigorous moderation process to ensure they do not contain copyrighted content, trademarks, characters, logos, or artist names.
However, it appears that AI images were not specifically excluded from Firefly's training, which would have been consistent with the company's advertising claims. Adobe's website states that Firefly respects the creative community by training only on licensed or freely available data.
Adobe is technically correct in stating that it trained Firefly only on material from its own stock database. But it is not fully transparent in hiding the fact that five percent of the selected training data from its stock database now comes from AI images with unclear legal and ethical backgrounds.
The case illustrates the complexity of the ethical and legal discussions surrounding AI-generated content that is indirectly integrated into AI training materials and for which there are only usage rights, not copyrights.
The legal implications of these practices remain largely unresolved. Midjourney itself has been criticized for the fact that its generator sometimes produces images that are almost verbatim copies of its training material.
Theoretically, Adobe's use of Midjourney images for AI training violates the startup's terms of service, which prohibit the use of images generated by its services to train AI models.