Why I hate the Anthropic ban (and you should too)

From Matt Lane, Senior Policy Counsel at Fight for the Future who is old enough to remember BBS

In the mid-90s, a then-young Electronic Frontier Foundation (EFF) won one of the most important legal victories in the history of the Internet.

A few years earlier, the US government had classified encryption software as munitions and heavily regulated this software’s export. In response, EFF took on the case of a young PhD student named Daniel Bernstein to challenge this law in a ruling that established code as speech. This core principle defined the early Internet and has enabled much of the weirder and wilder projects of today.

On June 9th, Anthropic released its much anticipated Fable 5 and Mythos 5 models to the general public. Three days later, the US government used the same argument it used to restrict encryption—classifying these models as dangerous munitions—to cut off Anthropic access to all foreign nationals, whether or not they reside in the US. Because Anthropic can’t determine who is a foreign national without invasive “know your customer” ID checks, they had to shut off access to these models entirely. The ban also means that foreign nationals at Anthropic can’t even use these models internally.

I’d argue that these two episodes bookend the Internet as we know it. What’s next will depend on what we are willing to fight for.

It’s common to classify the stages of Internet development in several “eras.” Web 1.0 was the DIY Internet, where everything was harder and more complicated but anyone could quite literally contribute infrastructure and carve out their own spaces. I experienced this myself in hosting or running a number of servers, including video game servers.

Web 2.0 is the current Internet, where Big Tech takes care of everything for us but we don’t own or control anything, and have to submit to a surveillance capitalist business model that wreaks havoc on our personal liberties. When anyone does come up with something interesting or new, it likely gets bought up and folded into a large corporation. Even Wordle wasn’t safe.

Web 3.0 is the proposition that we can use increased tech literacy and new decentralized technologies to get some of that Web 1.0 magic back, without the frustration of spending our evenings debugging. Or moderating fan forums.

The issue is that Web 3.0 hasn’t taken off yet. It’s hard to extricate yourself from the Google octopus (although we’ve found a huge amount of interest). And unfortunately, everyone hates one promising new technology that might actually help us break through to the ideals of Web 3.0: large language models.

It’s easy to understand why. The headline LLMs are controlled by mega-corporations and tech bros that are about as evil as you can get. They use up enormous amounts of resources and threaten our jobs. But it doesn’t have to be this way. Open source models are catching up, and researchers in places like China are showing you can build competitive lightweight models that come close to Big Tech capabilities using many fewer resources. There is no moat around this technology. In fact, you can run a large language model at home on a moderately priced gaming computer. Your phone might even be running smaller AI models locally (especially if you have an Apple phone). This means there could be a future in which you run an open source LLM locally or on a private server that can help you fill in the gaps towards a Big Tech free existence. Because let’s face it, much of what was great about Web 1.0 relied on people donating free labor to run a forum or a server or a website. We just can’t demand that from people anymore, but maybe AI tools can take make these tasks easier so long as they are open source, user controlled, and ethically created and managed.

The potential of AI is scaring people, some of whom believe that a misstep here could be an existential risk to humanity. That’s led to some mixed reactions to Trump’s extreme reaction to Anthropic’s new models. Some of the most respected leaders of digital spaces are wondering if maybe it’s a good idea to limit or ban frontier models, even while they are championing the vision for an open source and publicly controlled LLMs. And powerful venture capitalists like Marc Andreessen are giddy at the prospect of Trump hurting companies that compete with their own. Apparently, it is “based” to do a little crony capitalism. Meanwhile, much of the public doesn’t really know how to feel about this big mess.

This brings me to why I care about this ban, and why you should too.

First, let’s establish exactly what’s going on. Trump has no coherent plan or framework to regulate AI. Anthropic just happens to be on the outs with the administration. Anthropic isn’t evil in the way the administration wants them to be evil; they simply have ideological differences, hired a “radical democrat” as their security expert, and so on. The White House isn’t even right about how cyber-security works. They are caught up in the same bad thinking that criminalizes white hat hackers—that finding vulnerabilities is bad and keeping them hidden is good. Also, Amazon, which happens to be run by a billionaire who has restructured the Washington Post to suck up to Trump, just happened to take concerns about Mythos to Treasury Secretary Scott Bessent. I’m sure they don’t have a competing product in the pipeline.

Second, the authority the Trump administration is using can and will extend beyond Anthropic. Those Chinese open source models? Gone. The US hates China now. Just kidding! Xi and Trump just became besties and they are back, but OpenAI is gone now because Elon Musk hates them. Oh, and a new open source model is also gone because it keeps calling Elon a fascist. Weird, the new state sanctioned model loves Russia and keeps pressuring you to join the fight against “Antifa.” It also keeps snitching on you to the FBI. This also goes beyond just AI. What’s to stop Trump from classifying Signal or Tor as munitions?

Third, code is still speech. And if we reverse that, then say goodbye to any hopes of a decentralized Internet. Things like competition, free speech, the right to associate with people online, privacy, and personal security are constantly under attack by governments around the world. There are politicians that want you to upload an ID or scan your face to log on. They want to have a say in content moderation. They want to control what websites you can go to. They want to reward their rich tech bro buddies and punish people that disagree with them. All of these things are just barely being held at bay by the First Amendment and the fundamental belief that the freedom to speak and code are protected.

Protecting the Internet sometimes means we have to stand up for people we don’t like as long as what we’re doing is necessary to stop a takeover of the Internet by fascists and their corporate sponsors. The Internet is a Winchester Mystery House of good and bad ideas, positive and negative influences. The current iteration may suck, but we still have the power to strip it back to the studs and rebuild. But once its foundation gives way, buddy, we’re screwed.

Put more simply, we have the power to create better worlds on the Internet, to write and use the software we want. Actually doing so might be hard. It may even feel beyond us. But we have that power. We can’t let the government take it away from us by classifying software as “dangerous” and placing under direct government control. Everyone should be speaking out against this overreach by Trump.