According to reports, Adobe has been using images generated by competitors such as Midjourney's image AI to train its own Firefly image generator. This practice contradicts Adobe's previous claim that it only uses copyrighted or public domain content to train Firefly.
Previously, Adobe emphasized that its approach was different from competitors like Midjourney. However, according to these statements, AI-generated images (including Midjourney's images) accounted for 5% of Firefly's training data.
These AI images entered the training process of Firefly through Adobe's stock database. These images have been uploaded and sold on the platform for some time. In November last year, Adobe was criticized for not reviewing AI images related to the Israeli-Hamas war on its platform.
In response, a spokesperson for Adobe told Bloomberg that all images uploaded to Adobe Stock, including AI-generated images, undergo a rigorous review process to ensure they do not contain copyrighted content, trademarks, characters, logos, or artist names.
However, it seems that Adobe did not specifically exclude AI images from Firefly's training, which contradicts the company's promotional statements. On Adobe's official website, they claim that Firefly respects the creative community and only uses licensed or freely available data for training.
From a technical perspective, Adobe's claim that they only use materials from their own stock database to train Firefly is accurate. But they did not fully disclose a fact: now, 5% of the selected training data in their stock database comes from AI images with unclear legal and ethical backgrounds.
This case highlights the complexity of ethical and legal discussions surrounding AI-generated content. These contents are indirectly integrated into AI training materials and often only have usage rights rather than copyrights.
The legal implications of these practices are largely unknown. Midjourney itself has been criticized for its generator sometimes producing images that are almost identical to the training materials.
In theory, Adobe's use of Midjourney's images for AI training violates the company's terms of service. These terms explicitly prohibit using images generated by their services to train AI models.