- GitHub Copilot Enterprise and Google Gemini Code Assist are now being used for high-complexity tasks including OS and driver development, where code correctness and legal compliance carry greater stakes than in typical application work.
- The “Does v. GitHub” class-action lawsuit against GitHub and OpenAI alleges that Copilot reproduces licensed open-source code without meeting attribution requirements, a live legal risk for enterprises deploying AI code generation at scale.
- The U.S. Copyright Office’s position that purely AI-generated code cannot be copyrighted means enterprises must enforce meaningful human authorship in development workflows, or risk owning code with no enforceable IP protection.
AI code generation tools have moved well past autocomplete. GitHub Copilot Enterprise and Google‘s Gemini Code Assist are now being deployed on some of the most legally and technically sensitive work in software: operating system components, hardware drivers, and other low-level code where a licensing misstep can expose an entire product to legal challenge. The IP questions this raises are fundamentally different from the compensation debates in generative art, and in some ways harder to resolve.
A Different Kind of IP Problem
The creative-arts debate over AI centres on whether artists deserve compensation when their work trains a model. In software, the concern is more concrete. AI code models are trained on public and open-source repositories carrying licences, GPL, MIT and others, that impose specific obligations around attribution and redistribution. When a model reproduces or closely echoes a fragment of that code in its output, those licence obligations may follow the output into a commercial product.
This is the core of the “Does v. GitHub” class-action lawsuit filed against GitHub and OpenAIwhich alleges that Copilot reproduces licensed open-source code without meeting its licence terms. The case has not yet produced a definitive ruling, but it has forced enterprise legal and engineering teams to treat AI-generated code as a provenance risk, not just a quality one. For teams building OS components or drivers, where third-party code dependencies are common and licence obligations multiply, that risk is amplified.
The aesthetic question does not arise here. Functional correctness, hardware compatibility and clean IP ownership matter; whether the code has stylistic merit does not. That makes the legal exposure more tractable to measure, but no less real.
Ownership Gaps in AI-Generated Code
The U.S. Copyright Office’s position adds a second layer of complexity. Works created entirely by AI, without meaningful human authorship, are not eligible for copyright protection. For enterprises shipping products built partly on AI-generated code, this creates a genuine gap: code that is functional and deployable may carry no enforceable IP protection at all.
The practical consequence is that human involvement cannot be nominal. Developers who review, modify and make substantive decisions about AI-generated suggestions have a stronger claim to authorship than those who accept suggestions wholesale. How courts will draw that line in practice remains unsettled, but the risk of producing unprotectable code is real enough that legal teams at major firms are already treating it as a compliance issue, not a theoretical one.
This is a different problem from the generative art debate, where the central tension is between the model’s training sources and the rights of the people whose work informed them. In software, the output side of the equation is equally contested. Both ends of the pipeline carry legal exposure. For teams working on AI-assisted OS and driver development, understanding how tools like GitHub Copilot handle code review is one part of managing that exposure.
How Enterprises Are Responding
The response from major technology firms has been practical. Microsoftthrough GitHub Copilot Enterprise, and Google, through Gemini Code Assist, have both introduced enterprise-tier commitments around data isolation: private code used to customise models is not fed back into shared training data, according to both companies. Beyond platform-level protections, enterprises are putting internal governance in place, licence audits on AI-generated output, code provenance tracking, and clear policies on the level of human review required before AI-assisted code can be merged.
The emphasis on human oversight serves two purposes. It strengthens the authorship argument for copyright purposes, and it creates an audit trail that matters if licence compliance is ever challenged. Neither purpose is served by treating AI suggestions as automatically production-ready.
These measures add process overhead that teams accustomed to fast iteration may resist. The trade-off is real: tighter review cycles slow development, but they also reduce the exposure that comes from shipping code of uncertain provenance into products where IP ownership matters, which, at the OS and driver level, is almost always. Stay up to date with the latest AI developments at Auton AI News.
