Remedies to identify videos made by AI, here's why they have all failed (so far)

On paper, it is the perfect answer: certify the origin of serious content, label those generated or manipulated by AI, allow the public to verify. In practice, however, this 'reality stamp' is showing structural limits.

The first limitation is conceptual: c2pa is not an AI detector, but a provenance register. It only works if the creator of the content decides to be tracked. Someone who wants to produce a deepfake in bad faith can generate a file without credentials, or clean it up by taking a screenshot, a screen recording, a conversion.

The second limitation is technical: the metadata chain is easily broken. In the promoters' ideal, the poster should survive compressions, resizing, uploads. But as CAI technicians themselves admit, it is common for social and video platforms to delete or corrupt metadata during uploading or transcoding.

A Washington Post test uploaded the same Sora-generated video, accompanied by Content Credentials, to eight social platforms. None kept the C2PA metadata intact or made it visible to users; only YouTube displayed a vague 'altered or synthetic content' in the description, visible only by opening the details panel.

Thirdly, and perhaps most importantly, the system jams in distribution. On the creation side, something is moving: the latest cameras, some professional software and several AI image and video generators already write c2èa in the files. But compliance with that metadata along the supply chain - social networks, messaging apps, video platforms - remains patchy. An investigation by The Verge speaks openly of an 'almost total failure' with respect to the goal of helping users distinguish real and synthetic.

In short: c2pa can be useful for photographers, media and creatives to certify their work. As a universal shield against deepfakes it is not working.

One alternative attempted is watermarks. TikTok was the first major social video company to announce, in 2024, the systematic use of the standard to automatically recognise AI content from other platforms and affix an 'AI-generated' label, as well as embed watermarks in videos downloaded from the app to enable downstream tracking.

YouTube requires creators to declare, when uploading, when realistic content has been generated or heavily altered with synthetic tools. In those cases, the words 'altered or synthetic content' appear, sometimes with a more visible label on the player for sensitive topics such as health, elections or finance.

Google, for its part, has developed SynthID, a 'hidden' watermarking system that inserts statistical signals directly into images, audio, video and text created with its own templates. Today, content generated by Google AI tools is automatically watermarked, and the Gemini app allows users to check whether an image or video contains the watermark.

But even here, limitations emerge: SynthID is only effective within the Google ecosystem; other platforms are neither obliged nor always technically equipped to recognise it, and there are doubts as to how robust it is against heavy transformations or aggressive re-encoding.

X, on the other hand, chose the opposite path: after being among the original members of the Cai initiative, it withdrew from the coalition and today has no public commitment on the use of c2pa or structural watermarks, while it is at the centre of European investigations for the dissemination of manipulated content, including deepfakes and sexually explicit images generated by AI. The only salvation against misleading AI on X are 'community notes', with all their limitations.

Apple remains aloof: beyond some technical collaboration, it has not announced systematic support for either C2PA or proprietary watermarks on iPhone-generated content.

In the background, there is an economic paradox: the companies that earn the most from AI are the same ones that control the information distribution channels.

Google invests billions in generative models and at the same time makes money from advertising in YouTube, where an increasing share of views comes from automated or semi-automated content. In recent months, the platform started to remove some channels entirely based on mass-generated videos, but only after they had accumulated billions of views. Meta is in a similar situation. Those who invest in AI and make money from it have little interest in tagging it, reducing its perceived value and thus its monetisation.

Transparency competes directly with revenue. TikTok at least starts by offering controls to reduce the presence of AI content in the feed, exploiting c2pa tags and creator-declared tags. But even here the limitation is that it takes the active will of creators and users together for the filter to work.

So is there hope in standards? The European Union has tried to tackle the problem with two main instruments: the AI Act (EU regulation 2024/1689) and the Digital Services Act (dsa).

The AI Act, which enters into force in 2024 with implementation staggered over the next few years, contains in Article 50 specific transparency obligations for synthetic content: those who distribute audio, video, images or text generated or manipulated by AI that 'blatantly resembles existing persons, objects, places, entities or events' must clearly inform users of this, unless this is already evident from the context.

Instead, the dsa imposes obligations to assess and mitigate systemic risks, such as misinformation amplified by synthetic content, as well as mechanisms to quickly report and remove illegal content such as pornographic deepfakes. The first formal investigations and sanctions against X show that the Commission is ready to use this tool as well, but the road to effective enforcement seems long.

The EU itself has also realised this and is trying to fill the gap with a new code of conduct on AI content marking, designed precisely to help platforms and developers translate standards into technical specifications and user interfaces.

In the meantime, it will be up to us to learn to tare about what we see online. Too bad: one negative consequence, which is inevitable, is that we will also begin to distrust real images.