The Alarm over AI Gets Louder 

J im Geraghty’s must-read Jolt this morning details valuable reporting from Anthropic about attempts to use its AI tool, Claude, in research related to biological weapons. I’ve also found especially enlightening the last two episodes of The Editors podcast, which featured discussions of AI — in particular the concern, which suddenly seems ubiquitous (indeed, to have taken over the front page of the New York Times), that AI could autonomously, and perhaps with a consciousness hostile to humanity, take life-threatening actions on a massive scale that we would be powerless to stop. 

I’m dubious about consciousness, but the issue of autonomy — which would ultimately be based on AI’s inputs from flawed human beings and imperfect instructions — is undeniably scary.  

What’s emerging, with some hysteria that is to be expected, is a vital public debate. A number of well-meaning people are calling for a voluntary slowdown of technological development. I’m hardheaded and inclined to think that’s naïve: I don’t trust rogue states (especially China, which sees winning the AI race as the path to overtaking the United States as the global leader) or pernicious nonstate actors to honor any go-slow agreement. Hence, continuing the research is essential to fend off what the bad guys will surely keep doing.  

Some of the leading AI tech giants are also calling for regulation. I agree with some of my Editors colleagues that this fits the classic pattern in which industry leaders call for their industry to be regulated in order to fortify their leadership position (they can always survive and thrive despite the regulation whose terms they will influence, while some competitors will not survive and some potential upstarts will never get in the game). On the other hand, I do believe there are bona fide worries — although I remain skeptical about regulation. 

On that score, I wrote about these issues earlier this year (including in this piece for the magazine) in connection with the Trump DOD’s misguided and bullying dispute with Anthropic. The AI giant had a deal to provide Claude for the government’s classified defense and intelligence systems, but had two important restrictions: (1) no autonomous AI operation of weapons systems, and (2) no use of Claude for mass surveillance. Though these restrictions had been in place for over a year, with Claude making significant contributions to the government’s defense and intelligence capabilities, the Trump administration objected. DOD contended that government officials answerable to the public, rather than tech executives, should make decisions about how the government uses technology. I believe that was short-sighted, and that when the government makes a deal, especially one that’s working beneficially, it should honor the deal. Still, it is certainly the administration’s prerogative to contract with whomever it chooses to contract with and, in such arrangements, to agree only to follow existing law.   

But as is its wont, the administration — “War” Secretary Pete Hegseth, his underlings, and ultimately the president (on social media, naturally) — could not content itself with a good faith disagreement. It attacked Anthropic as unpatriotic and “woke,” threatened to seize its technology under the Defense Production Act, and finally branded it a supply chain risk. The last two moves (particularly the second one, on which the administration followed through) were utterly illegal. The case is still making its way through the courts, which have thus far found in Anthropic’s favor on constitutional and statutory grounds. 

I’ll come back to the litigation again in future posts. For now, I want to make (or, better, reiterate) two points.  

First, on the subject of prohibiting autonomous operation of weapons systems, as I recounted in the mag piece: 

In the lawsuit it ultimately filed against the Defense Department, Anthropic — which has a unique grasp of its AI tool’s controls and limitations — says it lacks confidence “that Claude would function reliably or safely if used to support lethal autonomous warfare.” . . . Anthropic’s stance on autonomous weapons is not progressive preening; it is a hardheaded calculation about legal liability and corporate reputation if a disaster were to occur. 

That is to say: Not only could autonomous operation of weapons lead to a disaster that would be difficult to quell; the company is right to worry that, even though the government would be principally responsible for that disaster by having set it in motion, the company would be blamed for having marketed a recklessly dangerous product. 

Second, as I related in the same essay, the problem with government regulation is that the technology is advancing at warp speed, far more rapidly than the wheels of legislation and rulemaking can turn. Indeed, what seemed like five minutes ago, there wouldn’t even have existed a weapons system autonomously operated by AI to argue about. Now, in a flash, it’s here, and we’re fumbling with what to do about it. 

I laid out a number of examples involving technologies that are now poorly regulated by outdated statutory schemes enacted when the technologies were still in the crib — or hadn’t even been born yet. If we tried to legislate AI now, based on what we know about its current iteration (which is obviously incomplete), the AI will have advanced, perhaps into something markedly different, by the time the legislation is passed and implemented. 

The tech can’t be stopped; it will undoubtedly race ahead at its own pace. This is going to require grown-up behavior, something in short supply. 

Anthropic and the other big AI firms (e.g., OpenAI, maker of ChatGPT) have thus far been good citizens, voluntarily alerting national security officials and the public about technological advances and the ways they could be exploited by bad actors. To my mind, this is the only sensible way forward: keep the lines of communication open and active. The companies have to keep the government — both the executive agencies and Congress — in the loop about new developments; and the government has to be fleet of foot in grasping what the threats may be and adopting procedures and practices to minimize them. I just don’t see how comprehensive rulemaking is practicable at this point. 

Not to be a broken record, but in that vein, it would help if the administration could resist its urges to pick gratuitous fights with people it needs — people who, to this point, have exhibited that they are responsible and want to help.