Can watermarks save us from deepfakes?

A video of Elizabeth Warren saying Republicans shouldn’t vote went viral in 2023. Nevertheless it wasn’t Warren. That video of Ron DeSantis wasn’t the Florida governor, both. And nope, Pope Francis was not sporting a white Balenciaga coat. 

Generative AI has made it simpler to create deepfakes and unfold them across the web. One of the widespread proposed options includes the thought of a watermark that will determine AI-generated content material. The Biden administration has made an enormous deal out of watermarks as a coverage resolution, even particularly mandating tech firms to search out methods to determine AI-generated content material. The president’s executive order on AI, launched in November, was constructed on commitments from AI builders to determine a technique to tag content material as AI generated. And it’s not simply coming from the White Home — legislators, too, are taking a look at enshrining watermarking necessities as legislation. 

Watermarking can’t be a panacea — for one factor, most programs merely don’t have the capability to tag textual content the best way it might tag visible media. Nonetheless, persons are acquainted sufficient with watermarks that the thought of watermarking an AI-generated picture feels pure. 

Just about everybody has seen a watermarked picture. Getty Photos, which distributes licensed images taken at occasions, makes use of a watermark so ubiquitous and so recognizable that it’s its own meta-meme. (In reality, the watermark is now the basis of Getty’s lawsuit in opposition to the AI-generation platform Midjourney, with Getty alleging that Midjourney will need to have taken its copyrighted content material because it generates the Getty watermark in its output.) After all, artists had been signing their works lengthy earlier than digital media and even the rise of pictures, in an effort to let individuals know who created the portray. However watermarking itself — in keeping with A History of Graphic Design —  started in the course of the Center Ages, when monks would change the thickness of the printing paper whereas it was moist and add their very own mark. Digital watermarking rose within the ‘90s as digital content material grew in reputation. Firms and governments started placing tags (hidden or in any other case) to make it simpler to trace possession, copyright, and authenticity. 

Watermarks will, as earlier than, nonetheless denote who owns and created the media that persons are taking a look at. However as a coverage resolution for the issue of deepfakes, this new wave of watermarks would, in essence, tag content material as both AI or human generated. Ample tagging from AI builders would, in concept, additionally present the provenance of AI-generated content material, thus moreover addressing the query of whether or not copyrighted materials was utilized in its creation. 

Tech firms have taken the Biden directive and are slowly releasing their AI watermarking options. Watermarking could appear easy, but it surely has one vital weak point: a watermark pasted on high of a picture or video might be simply eliminated by way of photograph or video modifying. The problem turns into, then, to make a watermark that Photoshop can’t erase. 

Firms like Adobe and Microsoft — members of the business group Coalition for Content material Provenance and Authenticity, or C2PA — have adopted Content material Credentials, a regular that provides options to photographs and movies of its provenance. Adobe has created a symbol for Content Credentials that will get embedded within the media; Microsoft has its personal model as effectively. Content material Credentials embeds sure metadata — like who made the picture and what program was used to create it — into the media; ideally, individuals will be capable to click on or faucet on the image to have a look at that metadata themselves. (Whether or not this image can persistently survive photograph modifying is but to be confirmed.) 

In the meantime, Google has stated it’s at the moment engaged on what it calls SynthID, a watermark that embeds itself into the pixels of a picture. SynthID is invisible to the human eye, however nonetheless detectable by way of a instrument. Digimarc, a software program firm that makes a speciality of digital watermarking, additionally has its personal AI watermarking characteristic; it provides a machine-readable image to a picture that stores copyright and ownership information in its metadata. 

All of those makes an attempt at watermarking look to both make the watermark unnoticable by the human eye or punt the exhausting work over to machine-readable metadata. It’s no marvel: this strategy is essentially the most surefire approach info might be saved with out it being eliminated, and encourages individuals to look nearer on the picture’s provenance. 

That’s all effectively and good if what you’re making an attempt to construct is a copyright detection system, however what does that imply for deepfakes, the place the issue is that fallible human eyes are being deceived? Watermarking places the burden on the buyer, counting on a person’s sense that one thing isn’t proper for info. However individuals typically don’t make it a behavior to test the provenance of something they see on-line. Even when a deepfake is tagged with telltale metadata, individuals will nonetheless fall for it — we’ve seen numerous occasions that when info will get fact-checked on-line, many individuals still refuse to believe the fact-checked info.

Specialists really feel a content tag is not enough to stop disinformation from reaching shoppers, so why would watermarking work in opposition to deepfakes?  

One of the best factor you’ll be able to say about watermarks, it appears, is that no less than it’s something in any respect. And because of the sheer scale of how a lot AI-generated content material might be rapidly and simply produced, a bit friction goes a good distance.

In spite of everything, there’s nothing unsuitable with the fundamental thought of watermarking. Seen watermarks sign authenticity and should encourage individuals to be extra skeptical of media with out it. And if a viewer does discover themselves inquisitive about authenticity, watermarks immediately present that info. 

Watermarking can’t be an ideal resolution for the explanations I’ve listed (and in addition to that, researchers have been in a position to break many of the watermarking systems on the market). Nevertheless it works in tandem with a rising wave of skepticism towards what individuals see on-line. I’ve to admit after I started scripting this, I’d believed that it’s simple to idiot individuals into believing actually good DALL-E 3 or Midjourney images had been made by people. Nonetheless, I noticed that discourse round AI artwork and deepfakes has seeped into the consciousness of many chronically on-line individuals. As an alternative of accepting journal covers or Instagram posts as genuine, there’s now an undercurrent of doubt. Social media customers repeatedly examine and name out manufacturers after they use AI. Take a look at how rapidly web sleuths referred to as out the opening credits of Secret Invasion and the AI-generated posters in True Detective. 

It’s nonetheless not a superb technique to depend on an individual’s skepticism, curiosity, or willingness to search out out if one thing is AI-generated. Watermarks can do good, however there needs to be one thing higher. Individuals are extra doubtful of content material, however we’re not totally there but. Sometime, we would discover a resolution that conveys one thing is made by AI with out hoping the viewer desires to search out out whether it is. 

For now, it’s greatest to be taught to acknowledge if a video isn’t actually of a politician.