The ongoing debate surrounding the training of large AI models has been a topic of contention since the emergence of the generative AI boom. One side of the argument is represented by tech companies such as OpenAI, who assert that training AI without resorting to copyrighted internet data is unachievable. On the other side, artists contend that AI companies have been exploiting their intellectual property without consent or compensation.
Adobe has adopted a distinctive position by aligning itself with the latter group, demonstrating that generative AI products can indeed be developed without the use of copyrighted internet data. The integration of the company’s image-generating model, Firefly, into the popular photo editing tool Photoshop, has served as evidence for this unconventional approach.
In an exclusive interview with MIT Technology Review, Adobe’s AI leaders underscored that it is indeed feasible to construct high-quality AI models without the exploitation of copyrighted content. This serves as a testament to the company’s dedication to ethical technology and equitable remuneration for creators.
David Wadhwani, senior vice president of Adobe’s digital media business, emphasizes the significance of pausing to consider the repercussions of developing new technology. He asserts that the mere possibility of building something does not justify doing so without taking into account the potential consequences. Adobe’s emphasis on the ‘how’ and ‘why’ has guided the development of Firefly, recognizing it as a stride towards producing responsible technology without compromising business opportunities.
As the generative image boom commenced, the company encountered significant opposition from creative communities. The utilization of generative AI models to produce derivative content and deepfakes engendered legal disputes over copyright and fair use. This compelled Adobe to take a stance that respects creators and provides legal certainty for businesses. They realized the crucial nature of constructing their models without scraping data from the internet.
Ely Greenfield, Adobe’s chief technology officer for digital media, expounds that Adobe sought to reap the benefits of generative AI while acknowledging that these advancements are built upon human labor. Consequently, they sought ways to equitably compensate individuals for their contributions. They trained Firefly on content with explicit licenses enabling AI training, primarily drawn from Adobe’s stock photos library. This diverges from the prevalent industry practice of indiscriminately scraping the internet for training data without a comprehensive understanding of its content.
Limiting Firefly’s training data to licensed content was a bold gamble, according to Greenfield, yet it ultimately proved successful. The decision furnishes Adobe with an advantage in content moderation and control over the images and texts generated by Firefly. Moreover, it ensures that the model refrains from generating objectionable or harmful content.
The company augmented the model’s training data with both human content moderators and automated content moderation. This facilitated the identification and filtration of objectionable or harmful content, as well as copyrighted material and images of recognizable individuals. It also furnished business customers with the ability to generate copyrighted content, fostering partnerships with companies such as IBM, Mattel, NVIDIA, and NASCAR.
Furthermore, Adobe has vocally advocated for labelling AI-generated content and has partaken in initiatives advocating for transparency and authenticity in AI-generated content. These endeavors serve to combat misinformation and uphold creator autonomy, establishing that investing in ethical AI has proven advantageous for Adobe’s business.
In summation, Adobe’s success with Firefly underscores the significance of responsible AI development and content moderation in the digital era, aligning their mission of empowering creators with the pursuit of societal welfare. Their ethical approach to AI has not only been beneficial for business but has also established a benchmark for the future of AI development.