As Adobe approaches the broader launch of its Firefly artificial intelligence offering, concerns have arisen regarding compensation for individuals contributing content to train their image-generation technology. One of the challenges Adobe is currently facing is the potential inclusion of AI-generated images in the dataset used to train Firefly. Adobe has emphasized that Firefly is more responsible than other AI applications as it solely trains on Adobe Stock images, openly licensed content, and public domain material. Additionally, the company plans to remunerate contributors if their uploaded content is utilized in training the Firefly model and generating new AI-powered images for Adobe customers. This situation raises complex questions that are currently troubling researchers and companies in the generative AI industry. Can one distinguish between AI and human-generated content?If not, what are the potential consequences?OpenAI, a leading AI company, recently confessed its inability to differentiate between the two. This suggests that AI-generated data might inadvertently be used to train new models, leading to a phenomenon called model collapse, where the quality of outputs steadily declines. For Adobe, if creators are uploading AI-generated images to Adobe Stock without disclosing this information, it could have a few concerning implications. Firstly, the company may end up compensating creators for content that was created by another AI model. Secondly, if the Adobe Stock platform becomes inundated with unmarked AI images, future iterations of the Firefly model might become compromised. Internal documents obtained by Insider indicate that Adobe appears less confident in distinguishing assets created by AI within its dataset. Adobe declined multiple requests for comment. The internal discussions and other documents reviewed by Insider indicated that employees and managers are genuinely grappling with complex AI topics. The final public announcements from the company may differ from the internal debates, but they shed light on how Adobe is developing this powerful technology. One question raised in the internal discussion document highlights the issue of identifying AI-generated content specifically. It asked whether it is probable that some contributors using generative AI would receive compensation due to Adobe's inability to identify all AI-produced assets. An Adobe manager instructed employees not to discuss mistakenly compensating contributors who have uploaded AI-generated images. Instead, they were encouraged to reiterate Adobe's previous statement that Firefly is not trained on images that are "known" to be AI-generated.

Another manager wrote that it would be best to avoid discussing these matters altogether. The manager stated that if a contributor who uploaded only AI-generated content receives payment mistakenly, Adobe should check if they accurately marked their assets, and if not, adjust the training model accordingly. "Let's handle these questions as they arise, " the manager suggested. Previously, Adobe mentioned its plans to develop a compensation model for Adobe Stock contributors and share more details later. The internal discussion also touched upon the contentious issue of using copyrighted or unapproved content to train AI models. Some of Adobe's competitors, such as Midjourney and Stability AI, are facing lawsuits for utilizing images obtained from the web without the consent of the original creators. Stack Overflow, a popular online forum for software coders, experienced a drop in traffic as users turned to competing AI models that scraped and trained on its freely available data, such as GPT-4 and Github Copilot. Adobe has an added urgency to address these concerns as an internal memo from July, reviewed by Insider, noted that the company was preparing for a wider commercial release of Firefly and anticipated general availability early in the fall. The Firefly beta was launched in March, and its generative AI capabilities have been gradually integrated into some of Adobe's other products. The memo mentioned that Adobe intended to update its Firefly FAQ for Stock contributors in September. As of late July, it stated that 1, 019, 274 Adobe Stock contributors were eligible for payment due to their images being used to train the Firefly AI model over the past 12 months. Approximately 6% of those contributors, around 67, 581, were expected to receive payments exceeding $10, according to the memo. Adobe is also contemplating allowing contributors to opt-out of having their images used for training the Firefly AI model. In the current Firefly FAQ, Adobe states that there is no opt-out option for training content submitted to Adobe Stock, but they are actively exploring the possibility. However, in a recent internal discussion, reviewed by Insider, an Adobe manager mentioned that contributors would soon have the ability to select "Do not train" when uploading images. If this option is chosen, the asset would be blocked from being uploaded to the Adobe Stock database, the manager explained. It remains uncertain if this will become Adobe's official policy. Insider reached out to the company for comment, but Adobe declined to respond. If you work at Adobe or have any tips, feel free to share.