The Trump administration wraps itself in the language of freedom, but reaches for the hammer of tyranny as soon as they see something screwy.
The present American administration freely uses the language of freedom. Free markets, free enterprise, the free world, “Freedom 250”. And yet, as soon as it sees something screwy, it reaches for a ban hammer.
The latest target is Anthropic’s frontier model, Mythos – or rather Fable 5, the publicly available, declawed version of Mythos – which has just been slapped with export controls so sweeping that the company has responded by blocking access for everyone.
Unable to reliably distinguish an American citizen from a foreign national across its global user base, Anthropic concluded that the only way to comply was to pull to plug on its most advanced large language model entirely.
Congratulations are due. In one stroke, Washington has managed to deny its own citizens and allies an advanced AI tool while doing very little to keep similar capabilities out of the hands of its greatest adversaries, or keeping its own technology systems safe.
This is, regrettably, becoming a pattern.
I have written before about the contradictory logic of export controls, and twice explained why China is emerging stronger from Trump’s trade wars than it went in.
The export ban on advanced AI is the same mistake in a new and even more self-defeating costume.
Dual-use
The ostensible justification is cybersecurity. Frontier models, we are told, are now capable enough at finding and exploiting software vulnerabilities that letting them loose abroad looks like a national security risk.
There is some truth in this. The most advanced models genuinely are formidable at this kind of work.
But this is precisely where the logic eats itself. Cybersecurity tools are, by their very nature, dual-use. The same capability that lets an attacker find a vulnerability is what lets a defender find and patch it first.
A model that can probe your network for weaknesses is the very model your security team needs to harden it.
You cannot hand the shield to the defender while denying the sword to the attacker, because they are forged from the same steel.
When an export control functions as a broad ban on the most capable tools, it does not disarm the attacker.
The attacker – a criminal syndicate, a hostile intelligence service, a freelance extortionist – was never going to fill in a licence application in the first place.
What the ban actually does is disarm the defender. It leaves individuals, public institutions, and companies exposed to exactly the class of exploit that the frontier models are best placed to discover and neutralise.
Substitute tools
Even on its own terms, the ban fails. The premise is that Mythos/Fable 5 possesses some unique, irreplaceable offensive capability. It does not.
Whatever cybersecurity research the model can do, a clutch of rival models can do nearly as well, and sometimes better. If there is a gap, it is measured in weeks and months, not years. Capability in this field is not something that can be guarded like the crown jewels. A whole fleet of AI models are steaming for the same frontier.
In February 2025, DeepSeek shocked the world by producing a language model that rivalled the best America had to offer, at a fraction of the cost, and in spite of AI chip export controls that were intended to prevent exactly that achievement.
The idea that a model from a jurisdiction hostile to American interests will not match or exceed Mythos’s capabilities, and soon, is not optimism. It is ignorance. It is wishful thinking of the most dangerous kind.
So the United States proposes to deny its friends a tool that its adversaries will shortly possess anyway, and in many cases already do. This is not security. It is a unilateral handicap.
It reminds me of the Cybercrimes Act in South Africa, a draft of which proposed to outlaw the possession of software tools that could be used for unlawful access, data interception, online fraud, and malicious communications.
Journalists and experts pointed out that not only are those tools already in the hands of everyone with a Linux distribution, but those tools are the essential means of defending against cybercrime.
At the time, I wrote: “Don’t outlaw crowbars and knives just because you can use them for burglary and assault.”
A chastened Department of Communications went back to the drawing board, and made the possession of such tools unlawful only when they are intended to be used to commit the crimes the Act prohibits.
Incitement to innovate
There is a deeper irony, which I have noted before. The administration justifies its import tariffs partly on the theory that making foreign goods scarcer will stimulate domestic production.
Whatever one thinks of that logic, it cuts directly against the export-control strategy. If restricting imports spurs a country to build its own, then restricting exports to a rival does precisely the same thing – for the rival.
Deny China, or anyone else, access to American frontier models, and you hand them the strongest possible incentive to fund, staff and build their own.
China graduates millions of STEM students a year, lavishly funds strategic research, and has already demonstrated that it will route around American restrictions rather than sit meekly counting ceiling tiles.
Export bans on software are not a wall. They are a starting gun.
This is what makes them futile against precisely the adversaries they are meant to stop. The well-resourced state actor has the talent, the capital and the motive to replicate the capability domestically.
The only people reliably stopped by an export control are law-abiding citizens – which is to say, the wrong people.
General-purpose folly
The root error is treating a general-purpose technology as though it were a munition. A large language model is not a guided missile with one terrible use.
It is closer to a printing press, or a programming language, or electricity: a foundational tool whose predominant application is benign and productive, and whose capacity for harm is just one facet of a larger whole.
To impose a sweeping ban on such a tool because it can be used for undesirable purposes is to throw out the baby, the bathwater, the bath and the plumbing besides.
It sacrifices the immense and overwhelmingly legitimate benefits of frontier AI – to medicine, to science, to programmers, to ordinary people going about their work – on the altar of a temporary, imagined safety.
Locking technology in a box
This should be a lesson for every government tempted to go down this road, including South Africa’s communications minister Solly Malatsi.
The instinct to regulate a powerful new technology by locking it in a box is as old as technology itself, and it has failed every time, from DVD encryption keys printed on T-shirts to the intrusion-software controls that the Wassenaar Arrangement signatories had to walk back once they realised they had criminalised the very security researchers they depended on for their online safety.
Software does not respect borders. Knowledge does not stay locked up. Information wants to be free (as in speech, not beer).
A government that grasps this will compete by out-innovating its rivals and equipping its own citizens to defend themselves.
A government that does not will see something screwy, reach for the hammer of tyranny, and hit its own thumb.
[Image: hammer2.webp]
[Caption: When all you have is a hammer, every problem looks like a nail. (Public domain image.)]