Making AI-Generated Content Easier to Identify
Before we begin, I have a confession to make: Truepic is not an AI company, and the Spaces we recently released are not demos of LLM models. So, who are we, and why are we here? Truepic is a leading provider of authenticity infrastructure for the internet, and we are here to create live demos that promote transparency for individuals using open-source models to generate images.
The emergence of high quality, AI-generated content has made it difficult to tell the difference between human and machine-created media. It becomes particularly crucial when an image shows something that never actually happened or tricks you into thinking it's real when it's not. Our collaboration with Hugging Face demonstrates how generative AI platforms can enable creators to label their content as computer-generated right from the moment of creation. We also worked with a forensic watermarking company Steg.AI to help showcase what may be possible, more on that below.
Truepic specializes in developing products for media software to cryptographically secure metadata into files at time of creation, editing, and publishing. We use the Coalition for Content Provenance and Authenticity’s open standard for metadata, referred to as Content Credentials. Truepic co-founded the standard body in 2021 with Adobe, Microsoft, Intel, and others. Unlike traditional metadata, Content Credentials are tamper-evident because they rely on hashing and are resistant to forgery because the data is signed into the file with an authenticated certificate. C2PA’s open standard can enhance the authenticity of EXIF data, label AI-generated or modified content, add authorship information, track edits, compare versions over time, and more.
Empowering Creators
Our first space, called GenAI with Content Credentials, lets you choose between a few text-to-image models hosted on Hugging Face and uses Truepic's technology to cryptographically secure metadata into every image you generate. This is done by programmatically adding Content Credentials to the file right after its generation using our command-line interface (CLI) installed on a server.
The credentials can then be viewed in an overlay on the image. You can download and transfer the media to any compliant platform, including editing tools like Adobe Photoshop, where your edit history can also be securely added to the file. This history travels with your media and can be extracted and displayed using a tool or website. Should you want to display on your own site, we also have a publicly available JavaScript library that can be placed on any page to elegantly verify and display Content Credentials. You can try it out in our demo. The display automatically shows an AI label when it appears in the metadata, allowing content consumers to easily identify it.
Experimenting for the Future
Our second space, Watermarked Content Credentials, is an experimental proof of concept. When an image is generated and signed, an imperceptible digital watermark is also added to the image pixels. This type of watermark is often referred to as an invisible QR code. To achieve this, we collaborated with Steg.AI, a forensic watermarking company and member of the C2PA, to integrate their service into our cryptographic signing process for the demo. The purpose of this watermark is to serve as a backup in case the Content Credentials are lost, such as when sharing the image between currently incompatible services, like text messaging. By using the watermark, it is possible to retrieve a restored, signed version of the image from before the data was decoupled. This approach can help improve the resilience of this critical information.
These demos represent a significant step forward in enhancing transparency and accountability in the digital world. We want to show that it’s not just about technological advancements; it's also about accountability, ethics, and creating a transparent digital future for everyone. Large tech companies have the resources to address these concerns, but what about smaller ones and individual developers? Hugging Face is an accessible platform, open to all, that prioritizes fostering a responsible environment, empowering diverse communities to assess AI's social implications, and guiding ML models' ethical development. We were thrilled to partner with Hugging Face because of their commitment to this mission.
We would love to hear from you, the community, if you would like to be able to sign metadata into the images generated here on Hugging Face. How can you let us know? Like our Spaces, and chat with us in the Community tab!
What’s next? Stay tuned for more exciting developments as we innovate and drive positive change. We plan to iterate and continuously demonstrate some of the more exciting methods through which we will enhance transparency, resiliency, and authenticity.