Maybe Figure Out What Happened Before Overreacting to the DeepSeek Bombshell?

N ational security officials and the stock market have been rocked this week by the news — we should more accurately describe it as the claim — that China’s DeepSeek artificial-intelligence model is performing on a par with American and other leading AI models despite its lack of access to advanced microchips. This development sent the stock of chip star Nvidia tumbling, although it has since begun to bounce back.

And maybe that’s because we don’t really know yet what happened.

Is it possible that China has developed the heretofore unknown technological prowess to train AI models on the cheap? Many knowledgeable people find that quite impossible to believe — notwithstanding initial reporting and market panic that framed this unlikelihood as established fact.

Far more likely is that China is cheating. Indeed, it is an incontestable fact that China is cheating by (a) hacking, (b) obtaining advanced microchip technology on the black market, (c) exploiting potential weaknesses in export controls, and (d) the process known as distillation in which, as the Wall Street Journal reports, China-based entities exfiltrate large volumes of data from openly available AI tools (e.g., ChatGPT, the OpenAI model) in order to train their own models. (That last one is a violation of Open AI’s terms of service, which seek to advance the technology by making it broadly available, but bar using it to evolve competitors. Memo to U.S. Big Tech: China doesn’t care about your terms of service.)

The question is not whether China is cheating. It’s whether there is a causative nexus between this array of illegal activity and the just-announced advance in DeepSeek’s capabilities.

If I were a national security official, my hypothesis would be yes. And that certainly seems to be the operating theory of David Sacks, the Trump administration’s AI “czar.” In the report linked above, Sacks told the Journal: “There’s substantial evidence that what DeepSeek did here is they distilled the knowledge out of Open AI’s models. And I don’t think Open AI is very happy about this.”

On the other hand, this sound theory should be pursued with modest confidence at best. We’re talking about dynamic (potentially life-altering) technology and global competition involving numerous players — good guys and bad guys. Lots of moving parts.

This is why I’ve been disturbed by some of the knee-jerk government reactions (at least the ones I’ve seen reported). Some congressional Democrats, for example, have leapt to the conclusion that DeepSeek’s claimed achievement makes it imperative to reenergize the Biden administration’s push for stifling government oversight: Compel AI companies to make disclosure to a regulatory agency (probably in the Commerce Department) when developing new models, establish federal development standards, further tighten export controls, and subsidize (i.e., pick winners and losers) the highly competitive chip industry.

In other words, in addressing challenges attendant to a technology that has developed in leaps and bounds through private companies and research — in connection with which it is in America’s vital interest to maintain our advantages over hostile regimes, particularly China, that are throwing mega resources into the competition — Democrats would have the government plop in with sclerotic industrial policy. (See, for instance, the policy that has modern telecommunications regulated under statutory schemes developed 30 or 90 years ago — before today’s modes and their underlying technology existed.)

Needless to say, we have been winning the AI competition thanks to the lack of government interference. Relatively uninhibited ingenuity and entrepreneurship is what enabled AI to emerge and evolve in the first place.

In another report, the Journal notes that the FBI has placed AI on a list of critical U.S. technologies. But it adds that “the rise of DeepSeek suggests rapid advances in AI can circumvent American know-how.” Well . . . not if China had to steal American know-how in order to achieve the rapid advances.

How about sticking to the things we already regulate and, you know, doing a better job?

For example, hacking is already against the law. Yet, China, Iran, and other countries are reportedly using AI to hone their cyber-attacking skills. So far, intelligence and law-enforcement officials assess that these and other regimes, as well as sophisticated hacking syndicates, are using the AI to do deep-dive research into potential targets — reconnaissance on how they operate, how their IT systems work, and the extent of their vulnerability to phishing and other known hacking tactics, etc. — rather than to create new cyber-attack methods. But that could be just around the corner. Are our investigators being aggressive enough?

Moreover, we already have export controls, but how effective have they been? For example, the U.S. has followed an “interdependence” strategy that, up until late 2023, enabled China to obtain semiconductors from Nvidia and other manufacturers . . . just not the most advanced ones. But the less-advanced chips were incorporated into other engineering processes that could close the performance gap, especially when complemented by better Nvidia chips on the black market.

Interdependence may be a sensible strategy: It keeps diplomatic channels open, and it discourages China both from more aggressive retaliation (on rare earth materials, for example, where it has the advantage) and from going its own way (which might leave us with sparse capability to monitor or influence its evolution).

But is interdependence the best strategy? That is, would it be better to cut off high-tech exports rather than try to modulate? I suspect that a cut-off would be better — and export controls have been ratcheted up in that direction — but I’m not competent enough that my suspicions are worth much. I do know this: It is a valid criticism that much of our policy in this area has been developed on the fly; it needs more strategic deliberation and careful implementation.

Bottom line: It’s not worth panicking over a day or two of surprise and careening markets due to DeepSeek. The ostensible breakthrough in Chinese AI may be more a function of mundane theft than suddenly emergent techno-wizardry. But we should get the facts before we jump to regulation, and we should be mindful that the usual suspects will exploit the opportunity to find a “crisis” that calls for more government control — control that would undermine the freedom that is America’s best competitive advantage.

Meantime, the government needs to be better at policing the wrongdoing it already has the authority to police.