OpenAI announced the development of new tools to detect content generated with artificial intelligence. To achieve this goal, the company directed by Sam Altman joined the management committee of C2PA (Coalition for Content Provenance and Authenticity) – a group created by Adobe, Microsoft And BBC to find methods of certifying content spread online – and has even started testing one internally new detection classifier technology that helps people detect material created with their own generative AI tools, first and foremost FROM-E 3, Sora And ChatGPT.
How the OpenAI system works to recognize AI content
With the adoption of the digital content certification standard – i.e C2PA – already used widely by several tech companies, online media and camera manufacturers, OpenAI can contribute to the development of this standard and use it to add C2PA metadata to all content generated with DALL·E 3 and ChatGPT. In the future, C2PA metadata will also mark content created with Sora, before the template text-to-video be launched on a large scale.
Let's be clear though: the C2PA standard may not prevent attackers from generating content without metadata. However, it is also true that these are not easily falsified or altered. Regarding this aspect, OpenAI has in fact reported:
As adoption of the standard increases, this information can accompany content through its lifecycle of sharing, editing and reuse. Over time, we believe this type of metadata will become something people will expect, filling a crucial gap in digital content authenticity practices.
In addition to adopting the standard in question, OpenAI is also internally developing new verification systems, including the implementation of tamper-proof watermarking. Thanks to similar technologies it is possible, for example, to mark digital content (such as audio) with an invisible signal that should be quite difficult to remove. Beyond that, the company is working on tools that use artificial intelligence itself to evaluate the likelihood that content is the result of generative models.
The internal tests conducted by OpenAI on its system proprietary detection classifier have highlighted a high precision in distinguishing images generated with DALL-E 3 from “real” images, achieving a accuracy level of approximately 98%.with an incidence of “false positives” (i.e. “real” images incorrectly marked as artificial) less than 0.5%. As regards checks on images artificially generated with third-party AI models, there is still a lot of work to be done, given that the system has achieved a rather poor level of precision, of 5-10%.
Why it is important to have tools to detect AI content
Maybe someone will ask why it is necessary to think about tools that detect content generated with AI. There are basically two reasons: the ever-increasing quantity of content generated by artificial intelligence models and the difficulty in distinguishing the latter from images, videos and texts created by “real” intelligence, i.e. human intelligence.
OpenAI itself, in its official press release, reports:
People around the world are embracing generative AI to create and edit images, video and audio in ways that enhance creativity, productivity and learning. As generated audiovisual content becomes more common, we believe it will be increasingly important for society as a whole to embrace new technologies and standards that help people understand the tools used to create the content they find online.
