The problem with calling models open or closed

Several major technology companies have urged Washington not to restrict open-weight AI models.

Anthropic is the notable exception.

Dario Amodei has clarified that he is not asking for a blanket ban. His concern is that sufficiently capable models may create risks that cannot be contained after their weights are released. Critics see this as protectionism from a company whose business depends on keeping its own models closed.

Both interpretations can be true.

I come from human genetics, a field that has spent decades arguing about openness. Much of modern genetics would not exist without shared software, reference datasets, and summary statistics. Researchers can build on work done by people they have never met, in countries they may never visit.

But this does not mean that every genetic dataset should be placed on the internet. Individual-level genomic data can reveal information not only about one person, but also about their relatives. Once released, it cannot meaningfully be recalled. We therefore share some things freely, place other things behind controlled access, and accept that the boundary will never be perfectly clean.

The useful distinction is not open versus closed. It is whether the benefits of access outweigh risks that may be difficult to reverse.

AI has the same problem, only at a much faster timescale.

An open-weight model can be inspected, adapted, and run without depending on a small number of companies. This matters for researchers, startups, and countries that do not want their technological future determined entirely by a few American firms.

But openness also changes the nature of control. An API can be monitored, modified, or shut down. Downloaded weights cannot. That difference becomes more important as models become more capable.

Anthropic is therefore not unreasonable to ask where the line should be drawn.

The problem is that a closed-model company benefits directly from drawing that line earlier. Even a sincere safety argument can protect a commercial moat. From the outside, principle and self-interest often look identical.

The only credible answer is to regulate capabilities rather than business models.

A sufficiently dangerous closed model should face the same scrutiny as an open one. A relatively low-risk open model should not be restricted merely because its weights can be downloaded. Evaluations, reporting requirements, and safeguards should follow what a system can actually do.

Anthropic’s position is principled only if the rules it supports would constrain Claude as seriously as they constrain its open-weight competitors.

Would they?

---