"Open-Weight" AI Models Aren't Open Source, and the Legal Gap Matters

Think about the last time you downloaded free software and trusted the license terms enough not to read them closely. That trust exists because decades of open source software history built clear, tested rules about who owns what and who's liable when something breaks. AI models don't have that history yet, and treating them as if they do could leave companies and individual developers holding legal risk they never agreed to take on.

That's the core warning in a new legal analysis from attorney Kate Downing, published through Practical Law. Her focus is on what the industry calls "open-weight" AI models: systems that companies, universities, nonprofit labs, and individual researchers post online for anyone to download. These models come with their "weights," the numerical values a model learns during training that essentially encode its behavior. Sharing weights lets anyone run or adapt the model without needing the massive computing resources required to train it from scratch.

The catch is in the name. Open-weight models get called "open source" constantly, in headlines, marketing copy, and casual conversation. But Downing's analysis is blunt about why that label misleads people: most open-weight models are closed in ways that matter a great deal to anyone trying to use them commercially or build products on top of them.

Open source software has a well-understood meaning, backed by license terms that have been tested in courts and refined over roughly three decades of use. Those licenses spell out, often in painstaking detail, what happens if code fails, who can sue whom, and how derivative works must be shared. AI weights typically arrive with none of that legal scaffolding, even when they're bundled with something called a "license."

Why the Old Rules Don't Fit

The mismatch runs deeper than terminology. Downing identifies several ways open-weight AI diverges from the open source playbook that companies have relied on for years.

Many of the licenses attached to open-weight models fail to actually protect the provider from liability, according to the analysis. A software company releasing free code typically disclaims warranties and limits its exposure through language that's been battle-tested in litigation. AI model providers often lack that same protection, meaning if their model causes harm or produces bad outputs, the liability shield they thought they had may not hold up.

The reverse problem exists too. Some of these licensing terms may be unenforceable against the people actually using the models. That's a mirror-image risk: a provider might believe certain restrictions apply, restrictions on commercial use, on modification, on redistribution, only to find a court won't enforce them because the underlying agreement doesn't meet basic legal requirements for a valid contract.

Layer onto that a regulatory environment that's shifting fast. AI is now subject to a growing patchwork of rules, varying by jurisdiction and by sector, that simply didn't exist when most open source licensing norms were established. A model provider operating comfortably under old open source assumptions could find themselves out of step with requirements that emerged after their model was already circulating widely online.

There's also a cultural gap. The people and organizations releasing open-weight models don't share the same norms or motivations as the open source software community that came before them. Some are companies looking to build developer goodwill or market share. Some are researchers publishing for academic credit. Some are hobbyists with no formal legal counsel at all. That diversity of motive means the licensing language attached to any given model can be inconsistent, aspirational, or simply drafted without full awareness of the legal consequences.

For a company deciding whether to build a product on top of a freely downloaded model, this creates real due diligence work. It's not enough to see the word "open" and assume the same protections and freedoms that apply to open source code. Companies need to understand both what the model can technically do and what the documentation actually promises, or fails to promise, in legal terms.

For the people releasing these models, the calculus runs the other direction. Downing's analysis suggests providers need to think clearly about what they're actually trying to achieve by sharing a model openly, and how much control they can realistically retain once it's out in the world. A license drafted with vague or borrowed language from the software world may not deliver the protection or the openness a provider intends.

This isn't a purely academic distinction. Open-weight models now power a huge range of downstream products, from chatbots to coding assistants to specialized research tools. Every one of those products inherits whatever legal uncertainty sits underneath the model it's built on. If the underlying licensing terms turn out to be unenforceable or fail to shield the original provider from liability, that uncertainty doesn't just disappear. It gets passed down the chain to every company and developer who built something on top.

Why It Matters

The stakes here go beyond legal technicalities for corporate counsel. As AI models become embedded in healthcare tools, educational platforms, and public services, the legal foundation underneath them needs to be solid. A mislabeled license or an unenforceable term isn't just a contract dispute waiting to happen, it's a gap that can leave real users unprotected when something goes wrong.

Companies evaluating open-weight models should treat the word "open" with healthy skepticism rather than as a shortcut past legal review. The open source movement earned its reputation for reliability through years of legal refinement. Open-weight AI hasn't had that runway yet, and pretending otherwise risks building critical infrastructure on ground that hasn't been tested.