Adobe has consistently promoted its artificial intelligence (AI) image generator Firefly as an ethical AI solution, primarily training it on a vast collection of licensed stock images sourced from the Adobe Stock library, comprising millions of images, designs, and illustrations.
Recent findings reported by Bloomberg suggest that the performance of this noise-cleaning photo model may not align with the idealized image projected by Adobe. There are concerns about potential inadequacies in the sourcing of training data for this “economically safe” AI model.
It has come to light that a portion of the images utilized to train the Firefly AI model were sourced from a competitor, Midjourney, without clear disclosure of the data origins. Speculations indicate that these images might have been sourced from the internet without proper authorization.
Adobe has acknowledged that approximately 5% of the extensive image dataset used to train the Light AI model falls into this category, although all images were sourced from the Adobe Stock collection, ensuring they underwent a stringent vetting process.
The significance of this issue lies in Adobe’s initial offering of indemnification against copyright infringement claims to entice enterprise clients, positioning Light as a secure alternative to competitors like Midjourney and DALL-E, provided that the training data was legitimate and licensed.
Despite the controversy surrounding the training data sources, Adobe maintains that all non-human images, including those potentially sourced from questionable origins, are protected under copyright law. Each image, whether sourced traditionally or generated through AI, undergoes thorough moderation to safeguard against intellectual property violations.
In light of these developments, Adobe seems to be expanding its AI ventures by venturing into AI video creation, reportedly compensating video contributors based on the duration of their content.
For further insights on this matter, attempts to reach out to Adobe for their official statement remain ongoing.
