I came across something on Engadget about AI watermarking being implemented to comply with new EU regulations. The topic was a new one for me, and piqued my interest. A lot of people have rather strong opinions on where and how artificial intelligence is used, and watermarking is a step to provide some visibility into that use.
What is AI watermarking?
As a general concept watermarks for AI are not dissimilar to watermarks for copyright. An image or other content is changed in some way to make the origin of the content apparent either visibly or cryptographically. Cryptographic watermarking hides the mark within the content so that it can be confirmed by suitable decryption and scanning software.
The European Union laws under something called the EU AI Act, specifically article 50 of that act, requires all products of generative AI to be marked in machine-readable form so that the source of the content can be verified. The law comes into effect for new AI systems as of August 2nd, and there is a grace period for pre-existing systems of at least several months.
There are a lot of good reasons for such an act that go beyond mere opinions: people can be harmed by AI generated content, for example via things such as “nudify” apps and political deepfake content generation. Being able to trace the origin of a piece of generated content seems like a good compromise solution between laws that prevent AI use and free, unfettered AI generation. Law abiding large scale companies like OpenAI and Anthropic will comply, but I’m skeptical about whether there will be full compliance amongst the most egregious deepfake image and video applications.
Techniques have been developed in recent years like SynthID that permit implementation of watermarks in things like application code. The ‘non-functional’ parts of the code and patterns within the code itself are not truly randomly generated in this model. A tool can be used to extract the watermark from application code that otherwise looks completely devoid of any such markings. This is the method that Anthropic is using, at least based on my simplified understanding of the matter.
My thoughts on the topic
Deepfake videos are often “good enough” to fool an average viewer into believing something that even surface level skepticism would clearly show is untrue. This is especially the case for fake videos and images that tell a story someone already wants to hear: manufactured outrage is easy when the audience are willing participants in the deception. So for the majority of outright fake content, AI watermarking won’t cause any improvement in public discourse.
Where AI watermarking will have an impact is with legal frameworks and governments that have associated protections against things like political fakery, factually incorrect news, and hate speech. The watermark will give plaintiffs and government bodies a tool to track down the source of illegal or offensive materials. This can then stand in a court of law to help prosecute otherwise unclear cases.
As I mentioned in the prior section, I’m somewhat skeptical about compliance with the law amongst the worst deepfake app vendors. The companies that produce these things likely already have little legal standing in the countries where the law applies. Nudify apps, for example, are probably already illegal in many jurisdictions so expecting the developers to add logic to watermark the videos or images they produce seems kind of improbable. But the existence of the EU AI Act may influence app stores to do things like require compliance with the act which could have a bit of n effect on the bottom line of such apps.
Watermarking will also assist those challenging copyright with the ability to prove that the supposedly protected content was actually machine generated and thus not necessarily subject to the same legal protections as something created by a human being. This probably applies most obviously to software generation things like Anthropic’s Claude.
Code generated with Claude as an example can now be detected within code produced by a software vendor, possibly calling into question any copyright they may choose to put on their product. It raises questions similar to copyleft licenses like GNU that can ‘infect’ an otherwise copyrightable codebase and unintentionally turn a commercial product into a public domain one.
At the end of the day, someone who wants to produce fake video, images, or ‘copyrighted’ code artifacts using “unannounced” AI generation will still be able to find a way to do it. But the legal framework of the EU AI Act at least encourages the largest and wealthiest companies like OpenAI, Anthropic, and Google to make their generated content detectable. I feel that this is a net good, and am happy to see it is being rolled out globally by at least some vendors.
Header image is AI generated using ChatGPT. The three fingered hand is more than a bit creepy…