While on the other side of the Atlantic, the authorities are considering severely penalizing deepfakes, OpenAI unveils its arsenal for detecting false content. Will these tools be effective enough to protect the 2024 US elections?
As the 2024 elections approach, the rise of deepfakes makes the issue of information integrity more complex every day. Disinformation is currently perceived as a real danger for populations and the emergence of new powerful tools like SORA from OpenAI doesn’t help matters.
The company, aware of the danger, has just presented a new tool which, a priori, would be capable of detecting content generated by AI. However, the initiative leaves some experts rather skeptical as to the real effectiveness of these measures.
Can we really stop deepfakes?
The AI Image Detector was designed specifically to identify images produced by DALL-E 3, its own image generator. The latter displays rather convincing results since its detection rate reaches around 98%. However, when it comes to other AI models, it runs out of steam; and not just a little.
Its success rate drops sharply and stagnates between 5 and 10%. A performance that borders on the ridiculous. Expert in sniffing out the productions of its creator, but almost blind to the works of other AI. Rather funny.
How does it work ? To authenticate images created or modified by DALL-E 3, the system relies on tamper-resistant metadata. These act as a digital certification, guaranteeing the verifiable origin of the content and limiting the risks of falsification. However, given that this mechanism is limited to images from OpenAI tools, effectively combating deepfakes with it may be complicated. Especially in a digital environment where many other AI generators proliferate.
A bandage on a wooden leg
OpenAI’s “anti-deepfakes” strategy rests on one pillar: the adoption of the C2PA (Coalition for Content Provenance and Authenticity) standard announced at the beginning of the year. The first could be compared to Nutriscore, a form of watermark label aimed at reinforcing the transparency and authenticity of AI-generated content. By integrating manipulation-resistant metadata, the C2PA standard makes it possible to trace the exact origin of content, thus increasing confidence in its authenticity. This metadata is supposed to guarantee increased transparency on the content creation process.
However, the C2PA standard is far from perfect. Because yes, supposedly resilient metadata is not completely resilient. It is always possible to falsify them or even delete them. This flaw could be exploited to create deceptive content that appears authentic. OpenAI itself recognizes these limitations; a reassuring situation.
OpenAI is trying to erect barriers against the potential abuses of AI, we cannot remove them. However, the effectiveness of the measures planned to ensure that the 2024 elections take place without a hitch still remains very uncertain. The race against time to secure the digital media space is far from won. Especially when we understand that the creators of the tools posing this threat are not yet capable of fully controlling them.
Source : Ars Technica
1