Banning Open Source AI Would Be A Mistake

Nathan Lambert and Kevin Xu, who writes the Interconnected newsletter, published this on June 19, 2026. It was written as an op-ed for a general, non-technical audience; the media outlets they pitched it to passed, so it ran on their own newsletters with a request to forward it to anyone on the fence. Lambert's open-source-ai-reading-list summarizes it as the case that open models foster education, innovation and competition, three core American values.

The occasion is a burst of activity in Washington: a newly signed executive order to review AI models, a congressional draft from Representatives Obernolte and Trahan to legislate AI further, reports that the government might take shares in frontier labs, and an action the previous week barring foreign nationals anywhere from accessing Anthropic's most advanced models. The authors' fear is that the next step, intentional or not, regulates or bans open source.

Three values

The op-ed defines open source loosely, as a process that lets technology be shared, built and distributed publicly. It opens with scale: over 90% of the world's software is built on open source, which has produced more than $8 trillion of economic value.

The education argument goes back to the free software movement, started at MIT in 1983, when teaching, research or even improving a printer meant dealing with companies like AT&T and Xerox. Every student in every university, community college and bootcamp now learns on tools that movement freed.

The innovation argument is that open source gives anyone free tools and a community to turn an idea into something real, without fear of a lawsuit or a bill. Some of that stays a hobby; some becomes Meta, whose first version of Facebook was built entirely on open source software.

The competition argument carries the most examples. Linux, which runs over 90% of cloud infrastructure, was the answer to the Windows monopoly, enough that Steve Ballmer called it "cancer". Android let a string of competing phone makers exist before the iPhone could take the market. Databases, self-driving and chip design get a mention.

Then the pivot: "Does AI change any of this? No." Anthropic and OpenAI are described as a duopoly concentrating power through closed models. Anthropic in particular is accused of flexing monopolistic muscle by reducing its best model's capability when someone uses it to improve their own model. Open weight models are the only counterweight available to startups, schools and enterprises.

Safety and China

The safety section is three sentences of argument. The authors grant that open models reaching frontier capability are worth monitoring, then argue that transparency makes open models safer, since more engineers can tune out unwanted behavior such as censorship and fix bugs in the serving software. Linus's law ("given enough eyeballs, all bugs are shallow") is quoted, along with Airbnb CEO Brian Chesky's point that a model running on your own infrastructure does not send data anywhere.

On China, the authors accept the rivalry but warn that using it to regulate open source will backfire. American startups in coding (Cursor's Composer 2 report is linked) and legal work already run on open models, including Chinese ones, because they cannot pay the premium Anthropic and OpenAI charge. That these models come from Chinese labs should be "a wake-up call that open source is under-invested and under-appreciated in America", which is the ATOM Project argument (atom-project-american-truly-open-models) in one line. Restricting open source because of China would chill education, innovation and competition at home and push the rest of the world toward Chinese models. The close borrows Louis Brandeis: sunlight is the best disinfectant, and open source is that sunlight.

Compared with Bedor

Tom Bedor's arguments-against-open-source-ai, written a month later, makes an overlapping case and the two are worth reading together. Both treat the history of open source software as the precedent, but they pick different history. Bedor's is encryption export controls, a direct case of the US government trying to suppress a technology and handicapping only its own citizens. Lambert and Xu use Linux, Android and the free software movement, which shows open source winning in markets but says nothing about what happens when a government restricts it.

Both see commercial interests behind open models. Bedor names Nvidia, American startups, enterprise buyers, and Google and Meta as constituencies with their own reasons to want open weights. The op-ed points at startups and Airbnb and paints Anthropic and OpenAI as the monopolists on the other side.

They part ways on China. Bedor argues that "losing the AI race" names no finish line, and that the solar-panel and steel analogies fail because software has no physical supply chain to capture. Lambert and Xu accept the competition and turn it into an argument for investing in American open models. That is consistent with Lambert's ATOM work, which itself leaned on worries about backdoors and censorship in Chinese models to justify building at home.

Neither piece engages the strongest safety argument: that a capable released model cannot be recalled, and that its capabilities, not its bugs, are the concern.

What it does not engage

The software analogy carries the whole op-ed, and it fits model weights less well than the authors suggest. Linus's law works for source code because people can read it. A set of weights cannot be read that way, and tuning out censorship shows that an open model can be modified, which is a different claim from showing it can be audited. The Chesky point is about privacy and data control, not safety. The op-ed also treats "open source AI" and "open weight models" as the same thing, which suits a general audience but blurs the distinction Lambert's own ATOM Project draws between the two.

A month later Lambert returned to the same fight in six-months-to-live-for-open-models, which links back to this op-ed as the reason a ban would be a mistake and adds the specifics missing here: which capability threshold might trigger action, how distillation became attached to it, and why he reads Anthropic's lobbying as regulatory capture.