On Monday, a policy analyst posted a video detailing a draft White House executive order that would require federal licenses for any AI model exceeding a certain capability threshold—effectively banning the distribution of open-weight models that rival GPT-5 or Claude Opus. The video, which ricocheted through developer Discords and Hacker News within hours, was treated as an act of war. The predictable response followed: calls to contact representatives, dark warnings about regulatory capture, and the familiar refrain that Washington doesn’t understand the technology it’s regulating.

That response is understandable. It is also a dodge.

The real state of open source AI, the thing the outrage conveniently skips past, is that the community has spent the last two years locked in a definitional knife fight it cannot win—and that paralysis is what made a ham-fisted executive order possible in the first place. The government didn’t wake up one morning and decide to strangle openness. It looked at a movement that couldn’t decide what “open” even means and concluded, reasonably or not, that someone else would have to decide for it.

The Definition That Ate Itself

The Open Source Initiative has been working on an Open Source AI Definition since 2024. The effort was supposed to bring clarity to a landscape where Meta calls Llama “open source” despite withholding training data, where Mistral releases weights under a research-only license and still gets the label, and where genuinely open projects compete for attention with marketing departments that have discovered the term’s halo effect.

Instead, the process became a proxy war. One camp insists that training data is the source code of AI, and any model released without it is a proprietary product wearing a costume. The other camp argues that full data release is impractical, legally fraught, and would hand a gift to bad actors. Both sides have a point. Neither side has been willing to budge.

In a Slack message to colleagues last month, one contributor to the OSI’s definition working group—speaking on condition of anonymity because the discussions are confidential—put it bluntly: “We’ve spent two years arguing about whether training data is ‘source code.’ Meanwhile, the White House is writing the rules for us.”

The result is a vacuum. And vacuums in policy don’t stay empty. They get filled by the people who show up with a pen.

The Convenient Outrage Cycle

Here is what the leaked draft actually does, according to the Monday video: it sets a compute threshold above which model distribution requires a federal license, with carve-outs for academic research and existing deployments. It is, in other words, a classic regulatory moat—one that would freeze the current competitive landscape in place. The companies that have already released frontier-scale open-weight models get grandfathered in. Everyone else gets a permission slip.

That should terrify the open source community. But the community’s response has been almost entirely performative. The same people who spent June arguing on GitHub issues about whether a model that releases weights but not training logs deserves the OSI seal of approval are now tweeting about the death of innovation. The connection between these two things—the infighting and the regulatory opening—is rarely made.

It is easier to rage against Washington than to admit that the movement’s own purity tests have made it politically illegible. When even sympathetic policymakers cannot get a straight answer to “what counts as open source AI,” they stop asking. They draft.

What a Real Defense Would Look Like

If the open source AI community wants to survive this moment, it needs to do something it has conspicuously failed to do: settle on a definition that is good enough, not perfect, and defend it with a unified voice.

That means accepting that training data transparency will be a spectrum, not a binary. It means acknowledging that a model released under Apache 2.0 with weights, code, and a detailed data card is meaningfully open, even if the raw training corpus cannot be redistributed for copyright reasons. It means building a coalition that includes the companies actually shipping open-weight models—Meta, Mistral, the Allen Institute—rather than excommunicating them for insufficient purity.

The alternative is to keep fighting about training data while the licensing regime arrives. And once it arrives, the definitional debate becomes moot, because the government will have supplied its own definition, and it will not be one the community likes.

Monday’s video may turn out to be a trial balloon, a draft that never sees the light of day. But the conditions that produced it are not going away. The state of open source AI, as the dashboard at stateofopensource.ai inadvertently illustrates, is a movement that has more energy for internal disputes than for the external threats those disputes invite. That is not a regulatory failure. It is a leadership failure. And the clock is ticking.

Sources