Cleanroom as a Service: AI-Washing Copyright

Earlier this month, Dylan Ayery and Mike Nolan shared a video of their presentation at the recent FOSDEM conference in Brussels.

Though much of the talk was an overview of the history of copyright law, it focused heavily on the idea/expression dichotomy and how copyright protects the expression of an idea, not the idea itself.

This is especially important in software as courts have previously found that it is legal to reverse engineer existing code so long as none of the original code is used. They specifically focused on the story of Phoenix Technologies, a company which produced the first IBM PC compatible BIOS by reverse engineering IBM’s offering.

At the end of the presentation, the duo announced the launch of their new service, Malus. According to them, Malus has the ability to “liberate open-source” by using AI robots to repeat the reverse engineering process performed by Phoenix Technologies.

The idea is remarkably simple. You have one group of bots analyze the requirements, documentation, and other public documentation of an open-source package. It then drafts a specification that is given to a second group of bots to generate the actual code. They claim that this produces “legally-distinct” code that can be used without complying with the open-source license.

To be clear, both the presentation and the site are tongue-in-cheek. They clearly intend this to be a joke and a criticism of copyright laws and the current state of programming. However, the service is operating and taking money from clients, as YouTuber The PrimeTime found out.

Whether it is a joke or not, it does raise some serious questions, and it’s not just computer programmers who need to pay attention.

The Big Idea

Under copyright law, there is a clear separation between the idea of a work and the expression of that idea. This takes many forms, but a good example is that the idea of a child wizard going to a magic school is not owned by anyone, but the Harry Potter series definitely is.

With software, there is a similar separation. For example, a function to check and ensure a user’s input is valid is an idea that no one can own, but the specific code to perform that function can be.

However, writing that code from scratch can be extremely difficult and time-consuming, especially on larger projects. You need clear standards and guidelines on what that the code has to do, and then write original code to do it.

Historically, this has been done with humans. When creating a code cleanroom, one group reviews the original code and documentation and then drafts a new standard from that. A second group, who has no contact with the first other than the standard, then writes the new code.

However, this can be time-consuming and expensive. At the very least you require two people and creating both the standard and the code take time and have to be done in sequence to ensure no contamination from the source code. As a result, it’s often easier to just use open source code and comply with the license requirements.

What Malus does is take the humans out of the equation. Now AI bots draft the specification, which is given to another bot to create the code. They claim that second bot is not trained on the original code, so it can’t and won’t produce infringing code.

However, before anyone gets too excited, there are some issues that those considering Malus’ approach may want to consider.

The Big Problem

The biggest problem is that I, nor would most programmers, trust AI code blindly. Trust in AI is at an all-time low and dropping further, this is a lot of trust to put into not one, but two separate AI systems to not just to produce effective code, but that it’s non-infringing.

Malus claims that the programming bots are not trained on the source material, but it doesn’t explain what LLMs its bots use. Their “Our Robots” link in the footer doesn’t work. It would be very difficult to train an AI programming bot without exposing it to open source software. In short, I would not take the (facetious) guarantees to heart.

But even if it did produce perfect code from a copyright standpoint, it’s not the only IP to worry about. Software patents can protect ideas and systems. Historically, software patents have been seen as antithetical to open source software, but it is possible to infringe on a software patent without copying any code.

But then there is the big lie, the Malus site says that, “The result is legally distinct code that you own outright.” That is completely untrue.

Under current US law, AI-generated works, including code, do not qualify for copyright protection. They are, essentially, public domain. While the company that used the tool would be free to use the code, they would not own it. Anyone else would be free to use the code for any purpose.

Even if the code is fully “legally distinct,” the “creator” would not own it. This may not be a big deal for most use cases, but if the goal is to create something that you can own and protect, AI generation is not the way to go, Malus or otherwise.

What About Other Works?

When it comes to copyright and patent law, computer code is in a bizarre place. As the 2021 Supreme Court ruling in the Google v. Oracle case highlighted, courts don’t look at computer code the same as literature, art and other kinds of protected work.

In that case, the Supreme Court ruled that the copying of APIs was fair use since the rest of the code was original. Though it’s a very fact-specific finding, it’s one that I could easily see a Malus (or other AI) user relying on.

But what about other kinds of work? What’s to stop someone from taking this approach and applying it to a book, a song or a movie?

To use the book example, it would be equally possible to have one AI system summarize the book and then use that as a prompt for a separate bot, essentially churning out a new version of the original work.

But that would be a problem. Though it would have to be judged on a case-by-case basis, it would be much more difficult for the second AI to produce something non-infringing.

The reason, it’s possible to create a derivative work of a book without copying any words. For example, fan fiction (usually) doesn’t copy any words from the source material. But it uses characters, locations and other details that are protectable. As such, works of fan fiction are unlicensed derivative works, even if they are more often than not tolerated.

With computer code, the protection is often limited to the code itself. If you avoid copying the code, you most likely avoid infringement. But that’s not as true with other kinds of work as other elements regularly enjoy protection.

Furthermore, if you are sued for creating an unauthorized derivative work, I can’t imagine the fact you used an AI cleanroom will do much to help your case. All it will do is prove the intent to create a new work based on the original.

Bottom Line

To be clear, I’m probably giving way too much time, energy, and oxygen to what is, ultimately, a joke. I recognize that. But as someone who is fascinated by copyright and authorship issues, this hits particularly close to home.

Obviously, if you were to use this tool, you would need to take all the normal precautions. You would need to check the code for completeness, ensure that it isn’t infringing and that you understand what it is doing. As someone who is already skeptical of AI, this approach seems to be even more risky.

In this article, I’ve deliberately avoided talking about ethics. Since Malus is presented as a tool for legal compliance, I wanted to examine it from that perspective.

That said, I think open-source software has been a major boon for programmers and users alike and those who build upon it should work to ensure that the commons is healthy. It’s the only way we can continue to benefit from it.

Ultimately though, I don’t think that copyright law is the problem. Though the idea/expression dichotomy has its issues, letting people lay claim to ideas would do far more harm than good. There’s a reason this has been a part of copyright law for hundreds of years.

There’s a long, difficult conversation to have surrounding AI and how it impacts copyright. However, upending a core tenet of copyright to address one particular issue is an overreaction and an overcorrection.

The biggest issue copyright has is how slow it moves when responding to new technology. Sometimes that is intentional and a positive thing, but sometimes, like with AI, it leaves a great deal of uncertainty and confusion as it is unable to answer important questions when those answers are needed the most.

There’s no easy fix for that. There may not be a fix at all.