The rollout of artificial intelligence technologies has landed at the worst possible time for humanity. (Well, the second-worst time; everyone having hallucinating robot friends during Covid would have probably been far worse.) At the same time, I’m a technologist who loves playing with all the new toys. So I was torn these past couple of weeks during the release and withdrawal of a powerful new Anthropic model, and the subsequent row between the Trump administration and the A.I. company.
On June 9, Anthropic publicly released Claude Fable 5, a more powerful model than its predecessors that was particularly good at writing and debugging computer code. After a few days, the Trump administration banned the model for any non-U.S. citizens, citing safety concerns.
Anthropic, lacking the ability to discern nationality among its users — and having been reportedly given only 90 minutes to comply with the administration’s order — pulled access for everyone. The government eventually lifted its restrictions, saying it was satisfied with steps Anthropic had taken to detect and address safety risks. By July 1, Fable was back online.
I am pro-A.I. regulation. So when I first heard that the Trump administration was cracking down on Anthropic, I thought: Even if this is happening in the most chaotic, least transparent way possible, perhaps it’s pointing in the right direction.
But over the Fable-less weeks, I found myself getting increasingly angry, not simply because I couldn’t gain access to the model I had paid to use, but also because it felt as if people were being censored, rather than protected. It agitated the same part of my brain that believes, quite deeply, in free speech.
And so what I’ve realized is that responsible A.I. governance may really come down to something resembling a First Amendment approach.
Before we get to that, take a step back. Do you vaguely remember hearing, back in April, about some big A.I. thing that could hack the planet? That was Mythos — the most powerful version of Claude, which hasn’t been publicly released. Anthropic discovered that Mythos could be pointed at large repositories of software code and would find very serious problems no one and nothing else had found, even in codebases that were decades old. Everyone was freaked out by the velocity and the realization that a bad actor doesn’t need to be a whiz kid to harness its power.
Those implications sent the global computer security status quo ablaze. Anthropic, trying to limit catastrophe, created a program for big companies, governments and open-source coding projects to get an early preview of the model so they could use it to find existing bugs and patch them up.
While it was figuring out how to deal with its new hacking superpowers, Anthropic added guardrails to Mythos that would put the brakes on when you asked it to, say, design a bioweapon or hack into a server. The new model that resulted from that, Fable, was pretty sweet: Before it was yanked, it found over 100 dumb security errors in a hobby project I’m building.
But Anthropic isn’t in the best of graces with this White House. (Remember how the Defense Department designated Anthropic a supply-chain risk when the company protested the possible use of Claude in autonomous weapons and domestic mass surveillance?) Many people assumed that the White House was out to get Anthropic and that its order to restrict the use of Fable by foreign nationals was motivated not strictly by national security concerns but also by antagonism toward the company. In contrast, OpenAI, America’s other major A.I. company, has reportedly discussed with the government giving it a 5 percent ownership stake — which is one way to keep everyone aligned.
Whether you take Anthropic’s side here or the White House’s, the events of the last several weeks are clearly a poor, disorganized way to approach governance of these models. I think they point to a troubling future. And that brings me back to the idea of free speech and code.
It’s easy to see large language models and chatbots as exhibiting pseudo-consciousness. By giving them chatty, subservient personalities and letting them refer to themselves as “I,” the A.I. companies have fed into the human tendency to anthropomorphize, causing immense confusion — especially when the bot is wrong, as it often is.
But these things are not conscious. They gathered vast troves of public and private information — sometimes without permission — and organized it into huge blobs of interconnected numbers.
Large language models, to me, are simply artifacts: objects made from the language expressed by tens of millions of people. And Anthropic, to me, is a software publisher. I do not see Claude as a living thing; I see it as a reference work prone to glitching, a surreal encyclopedia with wild side effects. I pay for the privilege of using it to generate new things — code mostly, and industry research reports.
But, critically, the things it generates are then mine, and I use them as I see fit. And, also critically, if I share them with the world, I bear responsibility for them. You see this with lawyers who submit A.I.-generated briefs — judges have no empathy for “the bot wrote it” arguments.
Prudence and slow-rolling releases might benefit the public in the same way that crash-testing cars is of benefit. Standards are good. But if the government can yank access at its whim, we’re starting down a slippery slope — the same slope we go down when we monitor library patrons and track what they read.
Now there are rumblings of ID-verification programs for A.I. users, to assure nationality. Do I need to be permitted to be an A.I. user, like I need to be to drive a car or own a gun? Do we want a future where Officer Claude pulls you over on the information superhighway and demands your license and registration?
A lot of work has gone into making A.I. safe — meaning, less biased and less harmful to vulnerable people (all while millions of dollars are pouring into lobbying the government on behalf of big tech and A.I. companies). But if industry trends continue, these models are going to fit nicely into our phones or desktops, and they might be as likely to come from China as they are to be from one of the big labs in the United States. If the U.S. government wants to keep regulating A.I. like it just did Fable, it will need something like the Great Firewall of China to block advanced A.I. models from being downloaded.
We’ve been here before. During the Clinton administration, the government did its best to lock down the export of advanced encryption technology. It didn’t want foreign entities to be able to keep secrets from America, though these actions had the potential to weaken online privacy everywhere.
Some hackers protested by printing crypto code on T-shirts or tattooing it on their bodies. Their argument was simple: You can’t censor math. Had these restrictions prevailed, the advancement of encryption might have suffered, making it hard to imagine how our global digital economy could work. You need provable security if you’re going to transact — that’s how you get Amazon, online banking, blockchains and Signal.
It is, I admit, appealing to imagine the government — well, maybe not this government — hitting the brakes on A.I. We could all use a rest. But ask yourself: Are you willing to compromise your free speech rights in order to keep people from finding bugs in your word processor? That’s what’s increasingly being asked of us when the government restricts the use of these models by certain groups, with little transparency and poorly articulated reasoning.
If A.I. is the new interface for creating code, and code is a form of communication and expression, then it’s incumbent to ensure only the most critical restrictions are applied to these models. And this is a new kind of technology, an industrial-strength symbol generator that often works in unpredictable ways. The balance is incredibly hard to strike, and it requires enlightened governance. Then again, there are established exceptions to free speech in America. We are capable of discernment.
Like all kinds of free expression, the right implies a responsibility. We should have a conversation about this that assumes the person who uses a system is capable of rational, civic action — versus applying a paternalistic view to these models and haphazard decision making over what is too dangerous to be used by the public. Especially because these systems were trained on public data. That would be more interesting than the conversation the Trump administration wants to have. I hope we can have it soon.
Paul Ford is an essayist and a technologist. He is a founder and the president of Aboard, an A.I.-powered software acceleration platform, and a co-host of the Aboard Podcast.
The Times is committed to publishing a diversity of letters to the editor. We’d like to hear what you think about this or any of our articles. Here are some tips. And here’s our email: letters@nytimes.com.
Follow the New York Times Opinion section on Facebook, Instagram, TikTok, Bluesky, WhatsApp and Threads.
