Introduction
Adobe’s image generation AI, “Firefly,” emerged in the market as a distinctive presence compared to other image generation AIs. Adobe emphasizes that Firefly is trained using a commercially safe library called Adobe Stock, claiming to clear any copyright issues. However, reports have surfaced indicating that Firefly’s dataset includes images generated by other image generation AIs. This blog delves into the discussions surrounding the commercial safety and transparency of Firefly.
Features and Safety of Adobe Firefly
Adobe Firefly is trained on materials that are free from copyright issues, such as images owned by Adobe or those in the public domain. Adobe highlights this aspect, asserting that Firefly is safe for commercial use. Furthermore, Adobe has declared that they will provide legal compensation if images generated by Firefly lead to litigation.
Issues with Firefly’s Training Data
However, according to a report by Bloomberg, Firefly’s dataset includes images generated by other image generation AIs, particularly Midjourney. This raises questions about the commercial safety claimed by Adobe. Image generation AIs like Midjourney, DALL-E, and Stable Diffusion train using images collected from the internet, which may involve copyright infringement and have led to class-action lawsuits.
Adobe Stock and AI-Generated Content
Adobe Stock began accepting AI-generated content at the end of 2022, and currently, 14% of the content is tagged as AI-generated. This includes images generated by Midjourney, raising concerns that these images are used in Firefly’s training. Brian Penny, a creator on Adobe Stock, expressed his surprise upon receiving a reward indicating that his work had been used for Firefly’s learning.
Challenges in Firefly’s Commercial Model and Transparency
It has been revealed that Adobe Firefly’s commercial model involves training on images partly generated by other image generation AIs. This contradicts Adobe’s claim of being a “commercially safe and transparent AI.” Raul Serron, Adobe Stock community lead, mentioned that a new training database excluding generative AI content would be used upon the official release, but the reality seems different.
Firefly’s Safety from a Legal Perspective
Rebecca Tushnet, a legal scholar at Harvard University, stated that even if Firefly was trained on image generation AI content, its safety from a copyright and trademark perspective would not be compromised. However, she pointed out that the inclusion of images generated by Midjourney contradicts Adobe’s stance of being different from other image generation AIs.
Conclusion
Adobe Firefly was developed with the aim of being a commercially safe and transparent AI. However, the inclusion of images generated by other image generation AIs has surfaced as an issue. This undermines trust in Firefly’s transparency and commercial safety, necessitating Adobe to ensure greater transparency and ethical practices. In the future, it is crucial to strictly select Firefly’s training data to restore the trust of creators and consumers.
This blog has discussed the challenges related to the commercial safety and transparency of Adobe Firefly and considered potential future responses. We welcome your comments and opinions.
