OpenAI to steer C2PA, label Sora videos as AI

One of several big announcements made by generative AI unicorn OpenAI today is that it is joining the “steering committee” of a trade group called the Coalition for Content Provenance and Authenticity (C2PA), founded back in February 2021 by Microsoft and Adobe (including Arm, BBC, Intel, and Truepic among its inaugural members).

Why? OpenAI says it is doing this to “help people verify the tools used for creating or editing many kinds of digital content” and to create “new technology that specifically helps people identify content created by our own tools.”

In other words: OpenAI wants to work with other companies in this space, including its rivals, to develop tools and tech for labeling AI-generated images, videos, and other content — allowing viewers to trace them back to their source and to avoid confusing them for real world footage and photographs.

What is C2PA and what does it do?

The C2PA organization, which operates under the non-profit Joint Development Foundation, is dedicated to “develop[ing] technical specifications for establishing content provenance and authenticity.”

In the three and a half-years since its launch, other big names in tech and AI including Google have also joined its steering committee, and the group has released a number of open source technical standards that developers and companies can implement on their products to make it clear where content generated by AI models or other tools comes from.

Among these standards are “the C2PA architecture; a model for storing and accessing cryptographically verifiable information whose trustworthiness can be assessed based on a defined trust model.”

The C2PA architecture has already been embraced and used by members of the organization to create “Content Credentials,” a web-friendly watermark indicated by a small “CR” icon in the top right corner of some images that users can hover their cursor over or tap to learn more information about who made it, using what tools, and when.

The C2PA architecture can also be baked into the metadata — or non-visual data that accompanies an image or video or other multimedia file — when it is saved, ensuring that even those who access it offline on other devices can see it. This is something OpenAI says it’s been doing for images generated with its DALL-E 3 image generation AI model since at least February of this year. (Meta has also begun labeling AI generated images with the C2PA standard.)

Now, today, Anna Makanju, OpenAI’s VP of Global Affairs, emphasized the importance of these efforts, stating in a press release from C2PA: “C2PA is playing an essential role in bringing the industry together to advance shared standards around digital provenance. We look forward to contributing to this effort and see it as an important part of building trust in the authenticity of what people see and hear online.”

Though it is not available to the public yet, Sora, OpenAI’s impressively realistic video generating AI model that is being used by selected trusted partners (including to make a first-of-its-kind music video last week), will also have C2PA metadata integrated to label video clips it generates as products of AI, when it is finally released to the public (no date given here).

In addition, OpenAI has launched something called the DALL-E Detection Classifier Researcher Access Program.

This initiative features a binary classifier designed to predict whether an image originates from OpenAI’s DALL-E 3 model. But the company wants help from outside groups in testing it.

Researchers interested in the program — OpenAI says the door is open to “research labs and research-oriented journalism nonprofits” — may submit applications until July 31, with decisions rolling out by August 31.

Social resilience fund

In addition, OpenAI says that with its investor Microsoft, it is launching a $2 million “societal resilience fund” that will partner with other external groups — including the AARP, International IDEA, and Partnership on AI — to “support AI education and understanding” among older adults and others unfamiliar with the tech.

The news comes amid reports of people, especially on social network sites such as Meta’s Facebook, apparently being tricked by posts featuring AI generated images meant to resemble real photographs — though many like “Shrimp Jesus” are quite obviously artistic and surreal.

The big question: will these efforts meaningfully help to curb the tide of AI disinformation? OpenAI is clearly attacking on both fronts — the generation and education component, so that AI generated content is labeled, but also that people learn how to recognize and look for it.

Yet it in an era when many open source AI models proliferate, and where it remains easy to screenshot or alter images to remove metadata (C2PA tries to make this difficult or impossible), it is clear that the challenge of reliably labeling and identifying AI content is likely to remain a formidable one for the foreseeable future. Still, the company wants to been as a good, social responsible force, so taking these steps makes sense from a public relations and ideally, good corporate citizen, standpoint.