Misalignment by design: How AI could become the operating system of the administrative state (Part 1)

The AI debate asks how the federal government should regulate commercial artificial intelligence and “big tech,” with little recognition or appreciation that the government has demonstrated itself to be ill-equipped to regulate conventional economic concerns effectively.

Contradictions abound. First the debate asks how Washington can make AI safe while ignoring the federal government’s own use, militarization, and expansion of unchecked AI capabilities through funding and political direction. It worries about job displacement even as government undermines employment and economic opportunity through spending, regulation, and intervention. The debate over AI professes concern over privacy and information collection while government itself prioritizes the likes of an attempt at mandating the reporting of $600 online payments received, the reauthorization of certain warrantless searches, and enactment of remote kill-switches for private vehicles. And then, it warns of concentration of economic power even as Washington increasingly pulls industries, investment, and economic decision-making toward itself like a black hole.

As it stands, AI risks becoming the technology through which government gains new means of controlling society, particularly if political institutions succeed in capturing the infrastructure upon which it depends, directing the development of AI science, and asserting certain redistribution authorities over the productivity gains AI generates. The urgent question, therefore, is not simply how the federal government should regulate AI, because by and large it should not. The question is how to prevent the governance machinery and political architecture now being constructed around artificial intelligence from taking root in the first place. Top-down “big science” is incompatible with limited government.

Federal ambitions for control are nothing if not conspicuous. The Office of Science and Technology Policy (OSTP) has issued new initiatives and funding priorities like the Genesis Mission, while OSTP Director Michael Kratsios and Office of Management and Budget Director Russell Vought have circulated guidance to agencies on Kratsios’s new report to the president called Science: A New Golden Age. These developments follow several earlier White House announcements and directives.

We have been here before, too, with aspirational AI governance having become something of a bipartisan fixture. The Obama administration issued a National Artificial Intelligence Research and Development Strategic Plan in 2016. Trump 1.0 brought its own federal AI initiatives and regulatory guidance.

Biden continued the dual display of technology-sector praise and enthusiasm coupled with familiar instruments of political direction: mobilizing agencies, coordinating research, subsidizing investment and development in service of a blueprint muscling in on competitive enterprise and federalism.

The separation between technology and state that a free society should jealously preserve is being steadily disregarded and cast aside. Explicit industrial policy is part of that development, exemplified in the new White House initiatives. Alongside it, AI is becoming a vehicle for a much larger political project: escalating the welfare state of the 20^th century rather than dismantling it and allowing the 21^st century to develop freer institutional foundations.

So the question is not simply, “How should we regulate AI?” The deeper necessity is to recognize these developments and ask what institutional arrangements are being coercively created around AI that will assure its political and economic capture and enable its exploitation in the service of dependency and political control. And then foreclose those as options.

The generic debate

Regulating AI as such repeats the familiar mistake of regulating technologies rather than addressing identifiable harms and conduct. Fraud, negligence, coercion, and other wrongful acts remain inexcusable whether or not artificial intelligence is involved, meaning no new all-purpose administrative apparatus for AI is necessary. Nor can preventing monopolization validate intervention when governments themselves are the source of coercive monopolization, exclusive franchises, entry barriers, and regulatory structures that prevent competition from emerging and sustaining itself in the first place. The conventional critique of top-down AI regulation has these concerns covered, but needs to dive deeper.

Control, not regulation, is sought

However benign or disciplined AI may prove in civilian life given the liability and institutional disciplines that ought to accompany powerful and risky technologies, the technology is simply too politically irresistible for the machinery of government and its allied interests to leave alone rather than exploit for political and cronyist gain. AI is increasingly embedded in infrastructure and objects, and regulating technology increasingly means regulating everything alluring. This misalignment is particularly troubling given the relative disregard in much of the AI regulatory debate for far more consequential military applications and for the enormous government and university research and funding apparatus surrounding artificial intelligence.

Here, too, there is considerable mythology surrounding supposed efforts to ban or stop data centers. The various political factions do not actually want to stop data centers. They want to control them, to affix AI infrastructure to new systems of governance that they can uniquely exploit. Mark our words: the last thing any power seeker wants is an independent, decentralized, and unhampered AI infrastructure beyond political reach. The current debate, therefore, with surprisingly broad political endorsement, is less about indifferent, arm’s-length regulation than about political control. Along with economic controls, this is headed toward the creation and management of an expanded welfare and entitlement state claimed necessary for “shared prosperity.”

The greatest danger is government becoming more powerful through AI

While Google’s one-time motto was “Don’t Be Evil,” a fitting admonition for today’s technology sector may be: “Don’t Be Government.” Isaac Asimov’s fictional but famous Laws of Robotics were designed to forbid harm to humans. The political danger today is that the institutional safeguards against government power are the ones getting “programmed” out. That is what makes the fusion of government and private AI deployment so problematic. The greatest danger may not be AI becoming too powerful, but government becoming more powerful through AI.

It is arguably true that we have been negligent in developing and defending appropriate liability regimes for genuinely harmful emerging technologies, particularly given those technologies’ longstanding capture by pre-existing regulatory regimes. Yet the dominant risk is not primarily AI doing bad things to us. The greater hazard is government using AI to do more things to us at a scale never before possible. Artificial intelligence dramatically reduces the cost of collecting information, processing it, monitoring people, classifying them, predicting behavior, and automating functions. In private hands and in voluntary transactions, those capabilities can be liberating. In the hands of government, they can become something else entirely.

A government that previously lacked the capacity to monitor, manage, classify, regulate, and otherwise interfere with millions of people is now acquiring that capacity. And once that capacity exists, the political pressure to use it will be immense. The great peril is therefore not simply that AI will replace human judgment or destroy jobs. It is that government will use AI to scale its own judgment, and its own frequent and widespread mistakes, beyond anything previously possible. That is why we should be especially suspicious of government claims of ensuring safety, privacy, and the like as rationales for controlling AI infrastructure and capabilities.

Read Part 2 tomorrow.