
In previous posts, I highlighted some of the real risks of artificial intelligence. Today, I’d like to talk about solutions. My PhD work is on an emerging solution that has been pioneered Dalhousie: Digital Locks for Machine Learning. I won’t be getting into the technical details of how they work but many people are familiar with digital locks from the era of digital rights management (DRM) and can remember trying to “rip” a DVD or “crack” a video game. You might even be reading this post on an “unlocked” phone, i.e. a phone that isn’t tied to a specific carrier.
Before I explain how this technology applies to artificial intelligence, let’s outline some properties of what makes a digital lock. First, the digital lock prevents some kind of unintended use: stealing a video game, copying a DVD, using a phone with a different carrier, or downloading a video from youtube or Netflix. Second, the digital lock, although not impossible to break, makes widespread unintended use much more difficult and expensive. Third, and possibly the most important, the digital lock is a critical piece of “regulatory infrastructure” or “governance technology’ i.e. the lock is a tool with which social policy and legislative mechanisms can be enacted.
How do these apply to artificial intelligence? Well, let’s look at another more popular safety solution: safety guards. Safety guards are the mechanisms in place that prevent ChatGPT from helping you with weapons development or the generation of deep fakes. So safety guards satisfy the first criteria. Additionally, the existence of safety guard mechanisms allows, for example, policy to be drafted that states that AI models cannot be publicly released without safety guards (meeting the third criteria). Unfortunately, safety guards do not meet the second criteria: they can be easily broken. Even more than easily broken, one major area of AI research focuses on improving training efficiency which means breaking safety guards will become easier over time.
There are many ways to break safety guards, but I will talk about the simplest and most effective. Machine learning or AI training is the process of learning from data using optimization algorithms. AI models are designed so that training is as cheap and effective as possible. This means that it very easy to collect (or create) a few samples of data of ideal behaviour (answer my unsafe question) without a safety guard and to train a model that is “uncensored.” It only takes a few cents to do this and datasets that undo safety guards are openly available on the internet.
So, when I talk about digital locks for artificial intelligence, I specifically mean mechanisms that prevent AI training or machine learning from working. This type of technology enables novel policy solutions such as the enforcement of digital locks on highly risky AI models or regulatory processes such as the automated certification that a digital lock is in place before a model is released to the public (or its continued automatic verification to enable usage). This technology also allows new legislative scenarios such as facilitating liability mechanisms by making model providers responsible for the outputs of their models so long as a digital lock was not tampered with and providing penalties for circumvention attempts or distribution of knowledge of how to circumvent as both are done with DVD copyright anti-circumvention legislation.
Unfortunately, like digital rights management, this technology could have negative effects like stifling commercial innovation (companies possibly can no longer train) or the enablement of anti-democratic technologies such as by locking highly censored models. Legislative mechanisms could result in aggressive lawsuits filed by companies who believe their models are being misused echoing cases such as Sony Computer Entertainment America, Inc. v. Hotz where George Hotz was sued by Sony for publicly unlocking the Playstation 3.
Despite this, there are a number of important differences between preventing training and preventing copyright infringement. Models can still be freely distributed and used for authorized purposes with a digital lock. Digital locks for AI could be constructed such that only training for illegal tasks are locked whilst other tasks can still be trained for. Digital locks for copyright primarily end up protecting profits since companies get to control which copies are made and sold by whom. AI has a fundamentally different economic model where the vast majority of the value comes from using the AI after it has been trained and therefore, unlike for copyright, digital locks don’t limit usage for personal or commercial purposes. Finally, the primary argument on the inevitable failure of digital copyright is that “bits can always be copied,” this isn’t necessarily the case for AI models as it isn’t clear whether “models can always be trained” given our recent findings at Dalhousie presented at the Atlantic Canada AI Summit 2025.
Image created with Imagen3 using the prompt “SNES 16-bit style robot with pad lock on the body”