Adobe has issued an update on its GenAI image product, Adobe Firefly. The company acknowledges that GenAI has caused a lot of uncertainty and concern in the creative industry regarding its use. This statement sets out how Adobe trains Firefly and protects copyright.
It goes on to say that it sees a role for AI as a tool for creators to expand their natural creativity. The technology is exposed through the Firefly website and used to power features inside its apps. The company also has a very active Discord channel where creators share what they are doing and how they get the most out of Firefly.
Interestingly, Adobe says, “We believe Adobe Firefly has the most creator-friendly approach in the industry.” It then addresses specific issues that are encountered by many AI systems.
What are the issues Adobe is addressing?
This document addresses two main issues: training and copyright protection. The full response can be found here. Many of the documents are linked, but the document also avoids something that many creators would like to avoid so that they can focus on their own work.
Training
The first issue is training, and Adobe addresses this in several ways. It makes it clear that it has never trained Firefly on customer content. Had it done that, it would have had to change its T&Cs. That would have risked people moving content elsewhere. That forms a key part of this statement, with Adobe saying it will only train on content with the relevant rights and permissions in place.
The two sources for training Firefly are Adobe Stock and public domain content, where Adobe says copyright has expired. For creators with content in Adobe Stock, the company has adjusted how it compensates creators.
With so much focus on AI training scraping the Internet, Adobe says it has not and will not do that. It says doing that is unfair to creators. Many other companies, such as X, Meta, OpenAI and Microsoft, do not share this view. They see publicly available content as free training data.
The company has also changed its terms and conditions to prevent third parties from scraping content to train their GenAI solutions. The T&Cs for sites such as Behance, where content is publicly hosted, explicitly state that content cannot be scraped. As we know from OpenAI, once content is in its GenAI solutions, it cannot be removed.
However, no document states what Adobe will do when content is scraped. Will it defend creators by launching legal action and seeking damages on their behalf?
Copyright protection
Adobe says it has multiple protections in place to respect creators’ rights. The statement reads, “We deploy safeguards at each step (before training, during generation, at prompt, and during output) to ensure Adobe Firefly does not create content that infringes copyright or intellectual property rights and is designed to be commercially safe.”
The company is also making no claim to any content created using Firefly. That is now part of its General Terms of Use.
Another part of protecting copyright is the Content Authenticity Initiative. There are now 3,300 companies signed up to it. It seeks to ensure transparency on who created content and how they did it.
Importantly, and this partly answers the question above over scraping, Adobe says, “it has taken and continues to pursue avenues for defending and protecting creator rights, including through proposing the FAIR (Federal Anti- Impersonation Right) Act.”
Enterprise Times: What does this mean?
This is an important statement from Adobe. It comes at a time when there is increasing concern over the impact of GenAI and image generation. It will be important to see if the other image generators follow suit. OpenAI is unlikely to, given its approach to scraping data from the Internet. The same is true of Microsoft’s Designer, which uses DALL-E.
While this lays out Adobe’s position, it does leave one area that creators want answering. When can they build their personal image libraries locally using Firefly? Many content creators and photographers have large image libraries that they have created over time. They might use Photoshop to edit that content but cannot create new content using Firefly from their personal content libraries.
What they want is for Adobe to extend Firefly so that they can build an image model based on their content. They want to then mix that content with the wider Firefly model. This is no different from how many organisations integrate local small language models with external content for GenAI across their business. This would give those creators much more flexibility and allow Adobe to develop Firefly and strengthen creator rights.
