Meta CTO Andrew Bosworth sees deepfakes less as a temporary confusion and more as a shift in how people judge media. In an interview with The Verge, he argued that watermarking alone is unlikely to solve the problem of unclear or fake AI content.
Why watermarking may not be enough
Bosworth does not believe in watermarking systems as a reliable answer for AI-generated media. His concern is rooted in a basic property of digital information: according to him, very few digital things cannot be reproduced digitally.
That creates a problem for both sides of the authenticity debate. A watermark meant to prove that something is real could be subverted. A watermark meant to show that something was generated by AI could also be subverted. In either case, the signal people are supposed to trust can become another surface for manipulation.
At the same time, Bosworth says Meta wants to take responsibility for its role on the Internet. The company’s latest image generator uses watermarking technology, so his position is not that labels have no place. The more limited point is that watermarking may not be strong enough to carry the whole burden of trust.
The unusual era of believable media
Bosworth frames today’s anxiety about deepfakes as part of a longer historical pattern. He says society has lived through a unique stretch in which photos and videos were almost certainly real. That period lasted for about the last 50 years.
Before that, people already had reasons to be cautious about written and oral accounts of events. In his view, visual information may now be moving into the same category. A photograph or video may still be useful, but it may no longer deserve automatic trust simply because it looks like a recording.
That is the core of his argument: the recent past may not be the permanent baseline. The last 50 years or so may turn out to be an exception, because it was strangely cheaper to capture real images than to generate artificial ones.
How people may adapt to deepfakes
Bosworth believes people will get used to fake media and adjust to the new reality. He compares the challenge with earlier problems society has faced and overcome. The implication is that skepticism may become a normal part of consuming all kinds of information, not just text or rumors.
"From today forward, it seems very likely that the youth of our world will know that all accounts that they come across, whether it be textual, oral, or video or photographic are suspect," says Bosworth.
That view does not make deepfakes harmless. It does, however, suggest that the central issue is not only technical detection. It is also public expectation: people may need to learn that realistic media can be generated, altered, or presented without the certainty that once surrounded photos and video.
Other AI leaders share doubts about detection
Bosworth is not alone in questioning authentication methods for AI content. OpenAI CEO Sam Altman is also skeptical of watermarking as a way to prove whether content is or is not AI-generated.
OpenAI recently took its AI text detector offline due to a lack of accuracy. That detail matters because it shows the difficulty of relying on automated systems to classify content with enough confidence for public trust.
Ian Goodfellow, the inventor of GAN technology that was the driving force behind the first deepfakes, made a nearly identical comment in 2017. He said the last few decades have been "a little bit of a fluke, historically," in terms of the authenticity of audiovisual information. In the future, people will need to be more skeptical.
What this means for AI-generated media
The debate around deepfakes is often framed as a search for a single fix. Bosworth’s argument points in a different direction. Watermarking can be part of a platform’s response, but it may not provide certainty in a medium where digital signals can be copied or undermined.
That leaves a harder cultural shift. AI-generated media may push photos, videos, speech, and text into the same trust category: useful, sometimes persuasive, but not automatically reliable. If Bosworth is right, the public will not only encounter more synthetic content. It will also become more practiced at questioning what any account is trying to prove.