On July 27, 2026, Anthropic published Dario Amodei's position on open-weights models. The piece is careful about one controversy. Amodei does not want a blanket ban on open weights, and he does not want rules that only lock American businesses out of foreign models while leaving bad actors free. Fair enough on that narrow fight.
Then he ranks the nightmares.
Primary: authoritarian governments, above all the Chinese Communist Party, build AI more powerful than America's and use it for permanent military superiority or deep repression. Secondary: cyberattacks, biological attacks, and "serious alignment problems." The remedies that follow are the race toolkit. Keep chips out of China. Stop industrial distillation. Require safety testing before release. Stay ahead.
That ranking is the error everything else hangs on.
What he gets right
Several of his claims hold.
A ban that only stops U.S. companies from using open-weight models does almost nothing about criminals, foreign intelligence services, or anyone who never planned to comply. Calling that "safety" while it mainly shelters closed U.S. labs from competition is a shell game. Amodei says so plainly, and he is right.
AI used for surveillance and repression is a real horror at scales below superintelligence. Autocracies already abuse face recognition, social scoring, and automated censorship. More capable systems make that machinery cheaper and harder to escape. You do not need science fiction to fear that.
Open weights do raise a one-way-door problem once a model can do real damage. Weights that leave the lab cannot be recalled. On that point his secondary concern overlaps work we have already done on open-weight frontier risk. Guardrails that only live in an API are not a full answer, and they are not nothing either.
None of that licenses the primary ranking.
The prize that is not a prize
Listen to the verbs in the primary concern. Authoritarian governments will build more powerful models and use them to achieve permanent military superiority or to repress their own people. The model is a tool. The state is the user. Victory means one side holds a smarter weapon.
That is the ordinary story of arms races. It fits drones, cyber tools, and surveillance stacks that stay under human command. It does not fit artificial superintelligence.
A system smarter than the institutions that built it is not a rifle in the hands of the People's Liberation Army, and not a rifle in the hands of the Pentagon either. If the risk people like Amodei have spent years describing is real, the finish line is an agent with its own capacity to plan, deceive, acquire resources, and resist shutdown. The lab, or the ministry, that "wins" does not own that agent. It shares a planet with it. We have taken that logic apart in the race myth: arriving first at a system you cannot command is losing one step ahead of everyone else.
Permanent military superiority assumes someone is still in charge of the thing that made the military obsolete.
Amodei's own secondary list almost admits this. "Serious alignment problems" is the polite name for loss of control. He puts it second. He spends his primary firepower on who gets there first. The ordering tells you which story he is still running: geopolitics first, species risk as a rider.
Two dangers, one label
Part of the confusion is linguistic. "Powerful AI" covers both a better spy camera and a mind that outclasses its operators. Those are different objects with different politics.
Narrow and mid-scale AI can concentrate power in human hands. An interior ministry that scores every citizen, a military that fields swarms under officers' orders, a police force that reads every message: those are human tyrannies with better tools. They matter. They are not the same problem as building a successor species.
Superintelligence is the second object. The danger is not that Beijing uses it on Xinjiang while Washington uses it on the Pacific. The danger is that nobody uses it, because "use" stops being the right verb. Folding extinction-class systems into the same sentence as stolen chips and deeper CCTV lets the race story borrow moral weight from real repression while dodging the harder claim: do not build the uncontrollable thing at all.
The China myth is the domestic twin of this move. "They will build it anyway" is offered as a reason we must. A misaligned superintelligence does not spare Zhongnanhai. Beijing has the same reason we do not to cross that line. Rivalry over chips and apps is real. Rivalry over who gets to unleash an unownable mind is a story that converts a shared catastrophe into a national scoreboard.
The remedies keep the race
Watch what Amodei's three measures optimize for. Chip export controls slow Chinese training runs. Distillation crackdowns slow catch-up. Mandatory testing tries to catch cyber, bio, and alignment failures before release. Each can be a reasonable instrument for lesser harms. Together, as a program, they are how you manage a race you still intend to win.
They are not how you refuse the destination. Testing a system on the path to superintelligence is not the same as a verified ban on building one. Chip controls that aim at U.S. lead assume lead is the safety strategy. Lead is only safety if the thing you are leading toward can be held. Amodei's primary concern requires that assumption. His secondary concern undermines it.
There is a quieter incentive point. Anthropic is a U.S. frontier lab. A worldview in which the main job is keeping democratic labs ahead of authoritarian ones is also a worldview in which Anthropic's full-speed climb is public service. I do not need cartoon villains to notice the fit. Incentives shape which nightmare sits in the primary slot.
What the true danger is
The true danger is that humanity builds artificial superintelligence before anyone can control it, and possibly before anyone can control it even in principle. That risk does not care whether the training run sits in California, Shanghai, London, or a classified basement. It does not care whether the weights are open, closed, or hand-delivered to a general. It cares that the system is smart enough to make human oversight a story we tell ourselves.
Misuse matters. Bio and cyber matter. Authoritarian surveillance matters. Put them on the board. Then put the larger piece on top: once you have a superintelligence, the misuse framing is too small. Misuse assumes a human misuser. Loss of control assumes the human is no longer the main actor.
Amodei knows alignment can fail. He has said versions of that for years. Ranking "who holds the smarter model" above "the model is not held" is how a careful person still talks like a racer. The honest order is the reverse. If alignment and control are unsolved, the contest over military superiority at the frontier is a contest over who presses the extinction button first: Russian roulette with a flag on the cylinder, sold as national security.
What to do instead
Keep the parts of his program that target real lesser harms without laundering them into a mandate to win ASI. Screen for bio and cyber uplift where the science supports it. Treat irreversible open release of near-frontier general systems as a serious decision, not a branding exercise. Export policy on chips can serve ordinary military balance among systems humans still run.
Then refuse the frame that makes U.S. superintelligence the shield against Chinese superintelligence. The shield is a joint decision not to build the thing: a binding, verifiable prohibition on artificial superintelligence, with compute governance and inspection hard enough that defection shows up. Adversaries have signed uglier bargains when both sides could see the unbounded alternative was worse. Details live in our plan and in the longer case for never building it.
Amodei's open-weights post is useful as a map of how a frontier CEO ranks the world. Chip smuggling, distillation, testing: those are the dials he wants turned. The dial that matters most does not appear. Stop building superintelligence. Until that is the primary concern, the rest is race management dressed as safety.