Anthropic, through Dario Amodei, clarifies that it isn't pushing to ban open weights and explains why it thinks a blanket prohibition isn't the right answer to AI risks. The text separates two main threats and suggests concrete measures to mitigate them, from chip controls to mandatory safety testing.
What Anthropic said and why it matters
In a public letter, Dario Amodei (CEO of Anthropic) rejects the idea that the company supports banning open weights. Why does that matter to you? Because the debate mixes legitimate safety concerns with political and commercial pressure.
Anthropic sums up its stance like this:
Open weightsmodels that don't have dangerous capabilities are a public good: cheap to use and useful for businesses, researchers, and developers.- Banning their use by companies in one country doesn't solve the most serious security problems. Many threats come from authoritarian actors or bad actors outside regulated spaces.
Sound familiar? In Latin America we've seen how open access can kickstart micro-entrepreneurship and local solutions. But we also know openness brings risks when technology scales fast without responsibility.
The two risks Anthropic worries about
Amodei follows two lines of concern he already explained in his essay "The Adolescence of Technology":
-
Main risk: authoritarian governments could build more powerful models and use them for military advantage or internal repression. Here the critical point isn't whether the weights are open: the real danger is that those models exist and are used secretly for control and armament.
-
Secondary risk: powerful models — open or closed — could enable cyberattacks, biological attacks, or alignment failures.
Open weightsmake it harder to apply safeguards and monitor uses because once released, they can't be pulled back.
But banning legitimate companies doesn't stop bad actors. That's the crux of the argument.
What Anthropic proposes (three clear measures)
Anthropic doesn't ask for blanket bans; it proposes targeted, concrete policies:
-
Don't sell advanced chips and manufacturing equipment to hostile actors. The idea is to cut access to the hardware that makes training frontier models possible.
-
Crack down on distillation at industrial scale. Distillation is a process that lets you create surprisingly powerful models with fewer resources by leveraging others. If used at scale with state support, it can undermine chip restrictions.
-
Require mandatory safety testing for all sufficiently capable models, whether
open weightsor closed. These tests should assess cyber, biological, and alignment risks before any release.
These measures aim at the root: who has the capacity to train and deploy dangerous models, and how we verify models don't pose real threats.
About the letter in favor of open weights
Anthropic recognizes the real values behind the letter promoting openness: access, competition, and customer control.
However, they disagree that openness necessarily makes safeguards easier or always benefits defenders over attackers. In domains like biology, the advantage can tilt toward the attacker; building robust defenses takes time and resources.
That's why Amodei insists the response must be empirical: test models before release and let the data guide regulatory choices.
What this implies for public policy and for you
-
Effective policies are specific: hardware controls, regulation of industrial practices (like distillation), and mandatory testing for advanced models.
-
We need international collaboration. An effective testing framework would have to be global to work: if one country opts out, the risk remains.
-
For entrepreneurs and developers: openness remains valuable, but you'll likely see stricter rules around high-capacity models. That will change how solutions are built and deployed.
As someone who writes from Maracaibo and has seen open tools empower local projects, I get the tension: how do you keep access and innovation without opening the door to serious harm? Anthropic's answer leans on targeted controls and systematic testing, rather than broad blocks that favor those who already dominate the market.
Final thought
Anthropic's note is an invitation not to fall for simplistic solutions. A total ban on open weights feels intuitive, but it's poorly targeted. Technical and policy measures — chip controls, curbing industrial distillation, and mandatory testing — attack the real causes of risk.
It's not easy or fast, but it's practical: protect society without completely sacrificing the benefits of access. Don't we prefer rules that work in the real world rather than bans that just change who competes?