Regulate What You Sow: Constitutional Duty and Existential AI Risk

The legal debate surrounding artificial intelligence focuses on what the government is allowed to regulate. Scholars and litigants alike have questioned the First Amendment’s power over algorithmic content moderation, whether artificial intelligence (AI)-powered surveillance is constitutionally viable, and whether Congress has delegated too much authority to agencies tasked with governing emerging technologies associated with the approaching AI age. While these are important areas for concern, they share a common assumption: the Constitution’s role in AI governance is restrictive, setting limits on government power rather than sourcing a government obligation. This article challenges that assumption, asking the more fundamental, unasked question of whether the Constitution doesn’t simply restrict potential actions but actually compels the government to act.

The case for this compulsion begins with the scale of the risk artificial intelligence poses. Geoffrey Hinton, the Nobel Prize winner often referred to as the “Godfather of AI,” resigned from Google in 2023 specifically to speak freely about his belief in the civilizational dangers AI poses [1]. As early as 2017, Elon Musk referenced the emerging technology as a “fundamental existential risk for human civilization.” By 2023, Turing Award-winning AI pioneer Yoshua Bengio urged the U.S. Senate that “we have a moral responsibility to mobilize our greatest minds” to protect “society and our shared future against potential perils” [2] [3]. Researcher and Machine Intelligence Research Institute co-founder Eliezer Yudkowsky has been more direct still, writing in 2023 that “The most likely result of building a superhumanly smart AI, under anything remotely like the current circumstances, is that literally everyone on Earth will die” [4]. While these experts differ in their tone and scale of casualty, they share an unmistakable underlying concern about the development of uncontrollable AI technologies, systems whose objectives we cannot predict or preempt. These warnings build towards a shared diagnosis: the existential risk posed by advanced AI is a structural, not intentional, one. The risk is not that the technology is or will be malicious, rather it is that the technology will be, and maybe already is, indifferent to human survival. Concerns of this magnitude question not only whether Congress should act, but whether the Constitution permits it to do nothing.

The constitutional right to life has already been proven capable of generating affirmative obligations in other domains. The emergence of existential AI risk demands that constitutional law consider whether that capacity extends to the technology with the potential to threaten human survival. Conventionally read as a negative right, a limit on what the government may do, not command to act, the Fifth and 14th Amendments combined/in conjunction/read together/in effect guarantee that no person shall be deprived of “life, liberty, or property, without due process of law.” [5] This reading is often used in limiting government power rather than sourcing government duty. This framing understates the Due Process Clause’s reach. In scenarios where the state is an active participant in creating a threat to life, courts have recognized that the Due Process Clause can impose affirmative obligations to act [6]. Environmental law has followed this arc exactly, moving from judicial refusal to constitutionalize ecological harm to the Supreme Court’s recognition in Massachusetts v. EPA (2007) that probabilistic, large-scale future risks can generate judicial duties [7]. In following the Due Process Clause’s positive application, DeShaney v. Winnebago County (1989), in which a state agency failed to protect a four-year-old boy from fatal abuse by his father, stands as the central doctrinal obstacle. The Supreme Court held that the Due Process Clause imposes no general duty on the government to protect its citizens from private harm [8]. DeShaney’s claim was dismissed under its own exception: the state-created danger doctrine, which imposes a constitutional duty precisely where the government has affirmatively acted to create or enhance a risk it then declines to mitigate. Artificial intelligence is a technology the federal government helped build, has formally acknowledged may threaten human survival, and has prevented states from independently regulating. It follows then, that the Constitution’s guarantee of life transforms that combination of facts into a constitutional obligation to regulate the existential AI risk.

Constitutional law has historically treated the Due Process Clause as a purely negative guarantee, however this framing obscures a critical distinction that determines when an affirmative duty can arise. Rooted in DeShaney and reinforced by decades of federal court precedent, the dominant reading holds that the Constitution is “a charter of negative liberties” telling the government what it cannot do, not what it must do [9]. Legal scholars have rightly noted the breadth of this principle [10]. Still, within it lies an underappreciated distinction between two categories of affirmative duty claims. The first category involves general welfare obligation referring to claims that the government must affirmatively provide services to such as healthcare, housing, or education. Consistently and correctly, courts rejected these claims under the Due Process Clause [11]. The second category involves state-created danger obligations referring to claims that the government must not create risks it then refuses to mitigate. Courts across multiple circuits have accepted these claims where the government’s own affirmative conduct enhanced the danger [12]. Put simply, the government cannot dig a pit, equip you with a blindfold, and then disclaim responsibility when you fall in. The constitutional question posed by AI risk falls squarely within this second category. It does not ask the government to provide a benefit, rather it asks the government not to fund, accelerate, and preempt regulation of a technology it has acknowledged may be existentially dangerous, while disclaiming any constitutional responsibility for the result. Importantly, even the post-Dobbs legal landscape does not foreclose this argument. While Dobbs narrowed the scope of substantive due process by rejecting novel social rights without roots in tradition or history, the majority’s disclaimer emphasized that “nothing in this opinion should be understood to cast doubt on precedents that do not concern abortion” [13]. The right not to be subjected to government-created risk is closer to the bodily integrity claims the Court left undisturbed rather than the kind of unenumerated social rights Dobbs targeted.

Life is not merely one interest among many in the Due Process Clauses. Life is the foundational guarantee: it is the first word and its protection pervades every other constitutional commitment. Unlike liberty or property whose boundaries have been subject to extensive doctrinal contention, the status of life as a Fifth and 14th Amendment-protected fundamental interest is rarely disputed [14]. The Supreme Court established a framework for identifying fundamental rights under substantive due process in Washington v. Glucksberg (1997). In the case, the Court rejected a claimed right to physician-assisted suicide and in doing so established a two-part test that now governs all substantive due process claims involving unenumerated rights. First, the test questions whether the asserted right is deeply rooted in the nation’s history and tradition. Next, it asks whether it is implicit in the concept of ordered liberty, the balance between individual freedom and societal welfare [15] [16]. The right to life clears both parameters trivially; no American law has ever endorsed the permissibility of government-facilitated existential risk. Glucksberg also requires that asserted rights be described with a high level of precision. The right being considered is not the abstract “right to life,” rather it is the precisely stated right not to be subjected to government-enabled existential risk. The magnitude of the threatened deprivation must also inform the constitutional analysis. The Supreme Court has recognized that the unique severity of a punishment demands heightened scrutiny. For example, the Court intensified its scrutiny of punishments precisely because the punishment was uniquely severe and grossly disproportionate in both Atkins v. Virginia (2002) and Roper v. Simmons (2005) [17] [18]. Existential risk extends this logic to the extreme. It is categorically different from ordinary threats to life in its totality and non-compensability. If severity and finality justifies heightened constitutional scrutiny in the context of a single defendant's sentence, they must justify at least equivalent scrutiny where the threatened deprivation encompasses the survival of the entire constitutional order itself.

The Ninth Amendment’s recognition of unenumerated rights provides a structural foundation that protects the right of the American people to persist as a self governing constitutional order, which is the condition for exercising every other right the Constitution guarantees [19]. In Griswold v. Connecticut (1965), Justice Goldberg’s concurrence established that, while not itself an independent source of rights, the Ninth Amendment “lends strong support” to the conclusion that fundamental rights extend beyond those specifically enumerated [20]. The claim that unenumerated rights do not exist is explicitly refuted in the Amendment’s text; it says that unenumerated rights do exist and that the people do retain them. Applied to existential risk, the argument here is structural rather than individual. A constitutional order presupposes the continued existence of the people who constitute it. Without the existence of the polity, any rights enumerated within the founding document cannot be exercised. The right to persist as a people is therefore a kind of meta-right: the precondition from which all enumerated rights derive their value. To protect rights, the Constitution protects the existence of the people who hold them. Without that inherent condition, the First Amendment protects no one’s speech, the Fourth Amendment guards no one’s home, and the 14th Amendment guarantees no one’s equal protection. The Court’s reasoning in Obergefell v. Hodges (2015) reinforces this structural reading. The Court held that “history and tradition guide and discipline this inquiry but do not set its outer boundaries,” signalling an openness to rights that define and expand expression of individual identity [21]. Physical survival is the necessary condition allowing these boundaries to exist. If the Constitution protects the right to define identity, it cannot be indifferent to the elimination of all identity. Thus, the Ninth Amendment functions as a structural reinforcement for the Fifth and 14th Amendment argument. It forecloses the objection that this right does not exist simply because the Constitution does not name it.

The central obstacle to any constitutional affirmative duty argument is Deshaney. In DeShaney, a four-year-old boy suffered permanent brain damage from repeated beatings from his father. Knowing about the abuse, state social workers still failed to remove the boy from his father’s custody. The Supreme Court held that the Due Process Clause did not impose an affirmative duty on the state to protect the child, reasoning that the Clause is “a limitation on the State’s power to act, not a guarantee of a certain minimal level of safety and security” [22]. Chief Justice Rehnquist’s reasoning rested on a bright-line distinction: the government’s failure to protect against private violence is not the same as the government’s infliction of harm. In the majority’s view, the Due Process Clause restrains only the latter. The Court did however recognize two circumstances in which affirmative duties come into effect. First is the custodial relationship exception: when the state takes a person into its custody and holds them against their will, it assumes an affirmative duty to provide for their safety. Citing Youngberg v. Romeo (1982), the Court held that involuntarily committed patients have a due process right to safe conditions [23]. Similarly, in Estelle v. Gambell (1976), they held that prisoners have a right to adequate medical care [24]. The second circumstance is the state-created danger exception. Lower courts have developed this exception from the DeShaney opinion’s reasoning: where the government does not merely fail to protect but affirmatively creates or enhances a danger, it cannot absolve itself of constitutional responsibility for the result. DeShaney bars the claim that the government must protect citizens from private harm. The relevant claim however is that the government itself built the danger. The claim is not that the government has failed to protect the public from a private danger and therefore must advance with positive duty. The claim is that the government has actively participated in the building, funding, and accelerating of the danger itself.

The state-created danger doctrine, though not yet recognized by the Supreme Court, has been adopted by nearly every federal circuit and operates through a defined multi-element test [25]. The Third Circuit’s example in Kneipp v. Tedder (1996) is the most widely cited. This test requires four elements: the harm ultimately caused was foreseeable and fairly direct, the state actor acted with willful disregard for the safety of the plaintiff there existed some relationship between the state and the plaintiff, and the state actor affirmatively used their authority in a way that created a danger or rendered the citizen more vulnerable than had the state not acted at all [26]. Other circuits have adopted similar tests. The Sixth Circuit requires that the state “created or increased the risk of harm” [27]. The Ninth Circuit asks whether the state “affirmatively placed the plaintiff in a position of danger” [28]. The Seventh Circuit asks whether the state ”put [a person] in a position of danger from private persons and then fail[ed] to protect him” [29]. While the Supreme Court has neither adopted nor rejected the doctrine, its silence does not undermine the argument. DeShaney created the analytical space for it by distinguishing between government inaction and government action that creates danger. The circuit courts adjudicated within that space. The Court’s repeated denial of petition on state-created danger cases, while not an endorsement, suggests no judicial hostility to the doctrine’s continued development and eventual application. The question now is not whether the state-created danger doctrine exists. Instead it is whether the federal government's relationship to artificial intelligence satisfies its conditions.

The federal government is not a passive observer of AI development; it is a decades-long funder, a formal acknowledger of the technology’s catastrophic risks, and a strategic accelerator of the very capabilities it now declines to regulate. The foundational technologies underlying modern artificial intelligence were developed with direct federal support. This support primarily came in the interest of national security. The Defense Advanced Research Projects Agency (DARPA) funded research that produced the core capabilities underlying today’s frontier AI systems including machine learning, natural language understanding, and neural networks [30]. Federal investment has only accelerated in recent years. The CHIPS and Science Act of 2022 committed over $280 billion to domestic semiconductor and AI research, explicitly designed to ensure American dominance in emergent artificial intelligence [31]. Diverging from its function as a regulator, the federal government is instead a co-investor with a financial and strategic stake in AI’s rapid advancement. While they continue to pump money towards AI acceleration, the federal government formally acknowledged that these technologies pose catastrophic risks. In November 2023, the United States signed the Bletchley Declaration, alongside 27 other nations, recognizing the “potential for serious, even catastrophic, harm… stemming from the most significant capabilities of these AI models” [32]. President Biden issued Executive Order 14110 that same year, directing federal agencies to evaluate risks from frontier AI systems and establishing the AI Safety Institute within the National Institute of Standards and Technology [33]. Beyond abstract acknowledgement, the government created the institutional infrastructure to address AI risk. In January 2025, President Trump revoked Biden’s AI framework and replaced it with an executive order declaring that the prior regulatory approach was “dangerous” and directing agencies to “revise or rescind” all policies that impede AI innovation [34]. A government that dismantles its own safety infrastructure is making an affirmative choice about how much risk the public will bear, exercising an authority over that danger and positioning it as an enabler instead of a passive bystander.

The federal government has also shaped the global conditions under which the AI risk develops. Since October 2022, the Bureau of Industry and Security has imposed sweeping export controls on advanced computing and semiconductor manufacturing products, limiting which nations can access the hardware necessary to develop frontier AI capabilities [35]. A government that controls the international distribution of AI-enabling technology, as well as willingly withdrawing proactive safety measures, cannot credibly claim it plays no role in the eventual risk that technology creates. These facts satisfy the state-created danger doctrine’s threshold requirement: affirmative government conduct that creates or enhances risk. The federal government funded AI’s development and accelerated its deployment all the while acknowledging its catastrophic dangers. Washington built and dismantled safety infrastructure while maintaining sovereign control over the technologies global distribution. This is a case of the government acting repeatedly, deliberately, and with knowledge of the consequences of the harm it was and is creating.

Constitutional doctrine has followed this path before: courts initially refused to constitutionalize environmental harm. Scholars then built a doctrinal bridge and the Supreme Court ultimately recognized that large-scale future harms can generate justiciable obligations in Massachusetts v. EPA (2007); AI risk is the clear next iteration in that evolution. For decades, courts rejected the claim that environmental degradation could invoke positive constitutional duties. Environmental harm was deemed too speculative, probabilistic, and controversial: all characteristics that made it a poor fit for the Due Process Clause’s traditional focus on individualized deprivations. In Tanner v. Armco Steel Corp (1972), the Court held that “no legally enforceable right to a healthful environment [was] guaranteed by the 14th Amendment or any provision of the Federal Constitution.” The majority opinion ruled that such claims were “inherently political” and “far too serious to relegate to the ad hoc process of government by lawsuit” [36]. Scholars then demonstrated that the state-created danger doctrine maps naturally onto environmental harms, reasoning that where the government affirmatively creates or exacerbates environmental risks, due process obligations follow. Shannon Roesler, a professor of environmental law at the University of Iowa, showed this exactly: that government participation in permitting environmental harm generates obligations that inaction cannot discharge [37]. Massachusetts marked the judicial breakthrough. The Court held that the EPA’s refusal to regulate greenhouse gas emissions was reviewable and the State of Massachusetts had standing to challenge it. They continued, ruling that a probabilistic, large-scale, future risk could generate a justiciable duty to act. The Court did not require proof that climate change would cause a specific, imminent injury to a specific plaintiff. Because the risk was real and the government had the authority to address it, its refusal to act was itself a policy choice subject to legal challenge [38]. Artificial intelligence presents a structural parallel. Like climate change and greenhouse gas emissions, existential AI risk is probabilistic, diffuse, and potentially catastrophic. And like the EPA’s refusal to regulate in Massachusetts v. EPA, the federal government’s current posture towards AI safety represents an affirmative regulatory choice, not inaction. Artificial intelligence necessitates the same doctrinal bridge. Further, the government’s participatory role in AI is even more direct than it is in most environmental cases from which obligations were derived. If constitutional law could evolve to recognize that probabilistic environmental risk generates positive obligations, it can recognize the same for a technology that leading researchers estimate poses a one-in-ten chance of existential catastrophe within the coming century. The doctrinal infrastructure for constitutionally obligated AI regulation exists, it just needs to be applied.

The challenges to a duty for AI risk-confronting regulation are clear. Although standing, ripeness, the political question doctrine, and separation of powers each pose genuine challenges, the Massachusetts framework answers all of them without requiring courts to step outside their institutional role. Massachusetts established that probabilistic, future-oriented harm is within jurisdiction of the court for standing purposes, extending “special solicitude” to states when challenging federal regulatory frameworks. As such, if a state attorney general challenged federal inaction over AI safety, they would be a sovereign entity alleging that federal regulatory abdication contributes to a large-scale risk that the government has acknowledged and still declines to address. This AI-example would occupy the same posture Massachusetts occupied [39]. Assessed under Baker v. Carr (1962), the political doctrine question presents a strong challenge with the “no judicially manageable standards” factor. However, courts are not meant to design AI policy. Instead, they should hold that complete regulatory absence in a domain where the government actively participates in risk-creation is constitutionally impermissible. This is the same institutional position the Court adopted in Massachusetts where it ordered the agency to make a reasoned finding without prescribing the substance of that finding [40]. Next, ripeness, whether a claim is too premature for judicial review, poses a legitimate concern. In Clapper v. Amnesty International (2013), the Court denied standing to plaintiffs who had alleged that the government’s surveillance program might target their communications in the future, requiring that the threatened harm be “certainly impending” rather than speculative [41]. Clapper is distinguishable by its facts: the plaintiffs there could not find evidence that the government had actually intercepted their communication or acknowledged any intent to do so. For AI risk, the Executive Order 14110 and the Bletchley Declaration clearly establish the government’s own conclusion that the risk is real. Government-acknowledged risk is not constitutionally speculative and avoids the ripeness challenge that killed the Clapper case. Finally, separation of powers concerns fail when the remedy is properly framed. Ideally, a court should hold that the constitutional duty exists and leave the specification of policy to the political branches; it is a declaratory judgement. Using the same model used in Massachusetts, the Court did not write emission regulations, it ordered the EPA to make a reasoned finding. This framework produces an institutionally modest remedy. While these structural obstacles are substantial, they are not insurmountable. Each has been overcome before in the environmental context. AI risk presents the next case for doing so.

The substantive objections to the constitutional duty to regulate existential AI risk are predictable, with each having a direct answer grounded in existing doctrine. First, DeShaney does not apply here. The ruling is limited to instances of government failure to protect against private harm where the state had no prior involvement. The federal government is disqualified from the DeShaney baseline because of its role in the direct funding of AI along with its acknowledgment of the existential AI risk. These factors place AI regulation squarely in DeShaney’s state-created danger exception instead. Next, the argument faces the same “policy dressed in constitutional clothing” charge every due process argument faces. The response from Griswold through Obergefell is that the Due Process Clause has substantive content or it is nothing more than procedure. The question is not whether the argument carries policy implications. Of course, all constitutional adjudication does. The question is whether the doctrinal grounding is sound. The third objection is that courts lack the technical expertise to properly assess and rule on AI risk. Courts regularly adjudicate claims that require substantial technical expertise: nuclear licensing, pharmaceutical approval, cybersecurity. Inquiring whether the government has abdicated a duty in a domain where it actively participates is a legal question, not a technical one. The fourth challenge focuses on the nature of existential AI risk reasoning: it is too speculative to ground constitutional obligations. The Bletchley Declaration and Executive Order 14110 overcome this objection on their own terms. Again, a government that has formally certified a danger exists cannot credibly claim that same danger is constitutionally speculative. Finally, the irreversibility of the threatened deprivation produces its own constitutional weight. The threshold for judicial engagement must be lower where the consequence of inaction is the permanent elimination of the constitutional order itself. These objections are serious, but none defeats the claim.

With three constitutional elements as its foundation, this article builds a framework by which to expand existing tools to previously unprecedented areas. The Fifth and 14th Amendments establish the protected interest: life. The state-created danger doctrine generates the government’s duty. The Ninth Amendment lays the structural foundation. Together, they answer a need constitutional law has failed to confront and cannot defer. Plainly, the government cannot fund frontier AI development, certify its threat as an existential danger, dismantle and reframe its own safety infrastructure and claim constitutional immunity from any duty to act all at the same time. Courts do not need to write innovative AI policy or revolutionary regulation. They only need to hold that complete regulatory abdication in a domain where the government actively participates is unconstitutional. Unlike most constitutional questions, the stakes of doctrinal delay here may lead to irreversible consequences. Materialized risk could result in a future without a court to reconsider the question posed in this article. There would be no plaintiff to see redress, and no constitutional order within which to assert any claim. That reality alone provides reason for courts to address AI risk promptly: the AI acceleration timeline by which humanity thrives simply will not allow for anything else. A judicially recognized constitutional duty would empower Congress to act with purpose and the vital injection of pace. That duty, grounded in the Constitution, would provide lawmakers with the authority to implement AI safety regulation under new constraints against inaction. The legal debate surrounding artificial intelligence has focused on what the government is permitted to do. Now is the time to focus on what the Constitution forbids it from failing to do, before it is too late.