The policy of President Donald Trump has, thus far, sided with innovation, particularly with respect to artificial intelligence (AI). But of late, a safety-obsessed insurgency has arisen within the White House. Weeks of public statements and counterstatements, leaks and speculation, have at least yielded an executive order, published on June 2, which directs the creation of a “voluntary” program to vet “covered frontier models” for the purpose of maintaining cybersecurity. The order’s import and implications do not immediately reveal themselves. Indeed, its chief virtues and chief vices both lie in what the administration chose to omit. If Joe Biden seized AI by the throat and sprinted off down the road to serfdom, the 47th President is stepping forward deliberately and cautiously. Whether Trump walks the same road as his predecessor or some other one remains unclear. The order’s aims are sound, but its implementation will determine whether it will be a boon for the AI industry or a regulatory barrage against innovation.
First, the executive order’s virtues: no licensing scheme for AI models will be imposed; the order explicitly prohibits it. The possibility seemed fanciful until last month, when Kevin Hassett, the director of the National Economic Council, stated: “We’re studying possibly an executive order to [determine] how future AIs that also potentially create vulnerabilities should go through a process so that…they’re released in the wild after they’ve been proven safe, just like [a Food and Drug Administration (FDA)–approved] drug.” The FDA, the epitome of safety-ism and bureaucratic excess within its own bailiwick, provides as suitable a model for regulating America’s most innovative industry as Wile E. Coyote provides for would-be pranksters. The agency’s operations are “slow and burdensome” (to borrow the President’s phrase); and its results, devastating to innovators within the industry it administers.
Having dispensed with Hassettian licensure, the Trump administration failed to dodge another pitfall: sheer vagueness. Security is, of course, a subjective concept, subject to debate. The executive order does not specify which models are to be “covered,” under what criteria models are to be evaluated, or how “to select [third-party] trusted partners” for the vetting process. Some of these lacunae are inevitable in a document directing agencies to act; but they are broad enough to provide bureaucrats extensive discretion in constructing—and, likely, remodeling—the architecture of the vetting process. Mutable standards, changed at the whim of regulators, would rob AI firms of certainty and subject them to the will of partisan and otherwise self-interested administrators. The safety-ist faction in the administration may have yielded today in the drafting of the executive order, but tomorrow and the next day will supply them—especially those in the national security bureaucracy, not distinguished for their attention to civil liberties—ample chances for mischief as it is implemented.
The “voluntary” nature of the vetting process is noble—and yet, any student of political economy knows that this may prove to be no more than aspirational, that such standards often prove all too compulsory. Guidance documents and bureaucratic suggestions too often carry within them a veiled command—as the experience of the financial-services industry, among others, shows. James Madison noted that “[a]n advisory government is a contradiction in terms.” Add thereto the increasing entanglement of leading AI firms and the federal government by means of government contracts, and one begins to foresee the assembly of a colossal incentive structure guiding innovators to comply for survival’s sake. Practically speaking, opting out may not for AI firms be a reasonable option. “Comply, or else” is not infrequently the information conveyed between the lines of “voluntary” regulatory provisions. And this administration has already revealed its taste for retribution against non-pliant firms—beyond any legitimate state interest—in its pursuit of Anthropic.
In the implementation of the executive order, the administration must decide what kind of technology sector it hopes to foster—one for which placating the demands of Washington, D.C., is the first thought, or one whose primary task is innovation. Regulatory regimes resemble molds: by shaping the incentives of the regulated, they shape the character of the regulated industry.
The other silence to be considered resounds from Capitol Hill, where, despite the multiple proposals introduced to regulate AI, lawmakers have thus far failed to enact a national standard. Although the White House might chafe at the legislature’s slowness to enact such a statute, congressional lethargy should not prompt the White House to reach for its pen and phone. Congressional action is not intended to be quick and easy; the legislative process of deliberation, debate, and negotiation requires time and effort. Basic, honest, and truly voluntary collaboration with innovators to stave off cybersecurity risks inherent in advanced AI is a worthy policy for the President to pursue, but he ought not succumb to the calls of the safety-ists within his own administration. The executive order must not, in its implementation, be transformed into a de facto regulatory measure.
