Ungoverned: AI, Accountability, and the Limits of Law
xCULR is an initiative by the Columbia Undergraduate Law Review’s Digital Initiative division that brings together undergraduate law reviews from institutions around the world in collaborative legal scholarship. This piece was produced in partnership with the Yale Undergraduate Law Journal (YULJ).
Section I: Federalism in AI Regulation, A Comparative Analysis of Trump and Biden's Legislative Approaches
The question of who has the legal authority to regulate artificial intelligence (AI) has become one of the most contested federalism disputes in American law. Under the Biden administration, the White House Office of Science and Technology Policy published “Blueprint for an AI Bill of Rights: Making Automated Systems Work for the American People” (hereinafter, “AI Bill of Rights”) in October 2022, a suggestion to develop AI policy through a civil rights framework. While the blueprint sought to be comprehensive, it was only a non-binding framework rather than an executive order or law. Without congressional legislation, the executive branch cannot create enforceable obligations on private actors, and the blueprint’s aspirational status left it vulnerable to removal by a subsequent administration, which is exactly what happened. Due to the non-mandatory nature of the blueprint, the Trump administration was able to remove the document from the inter-administration White House website following Trump’s election. Where Biden’s framework aimed to monitor AI’s integration into society through non-binding recommendations, Trump prioritizes AI’s uninhibited expansion by scrutinizing its legal barriers. This legal tension between the two administrations is situated in a system of federalist struggle: while the Biden administration left room for actors to interpret the policy framework without legal consequences, the Trump administration rejected governments’ ability to regulate AI at any level by inscribing the free-rein stance into legislation.
Despite invoking the gravity of the Bill of Rights, the “AI Bill of Rights” is non-binding and lacks any enforcement mechanism, and the legal consequences of that status extended beyond the blueprint’s eventual removal. During the Biden era, private actors who developed or deployed biased AI systems faced no legal liability under the blueprint because it created an expectation of responsible AI development without any corresponding legal obligation to deliver it. The blueprint claimed that AI can “reflect and reproduce existing unwanted inequities or embed new harmful bias and discrimination,” but provided no statutory basis for holding anyone accountable when that happened. This gap demonstrates a structural limitation of executive power: without congressional legislation, the executive branch can identify harms but cannot compel private actors to prevent them.
Biden’s Executive Order 14410: Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence attempted to go further but encountered the same structural limitation. Though the Order functioned as law within the executive branch, it did not create any new regulatory obligations for private actors or state governments. This is a recurring tension in administrative law: executive orders bind the executive branch but cannot impose obligations beyond it without statutory authorization from Congress. The Order outlined explicit social inequities that AI could exacerbate — “fraud, discrimination, bias, and disinformation; displace and disempower workers; stifle competition; and pose risks to national security” — and highlighted “advancing a coordinated, Federal Government-wide approach” as the most effective method of regulation. But without accompanying congressional legislation, the Order functioned as a directive to federal agencies rather than a regulatory framework with external legal force. Biden identified the right problem but lacked the constitutional authority to solve it unilaterally.
Trump’s approach represents a fundamentally different use of executive power. Rather than encouraging regulation, Trump has used executive orders to actively dismantle it. Executive Order 14179: Removing Barriers to American Leadership in Artificial Intelligence, aimed at funding and sustaining AI innovation and global competitiveness, identified the biggest obstacle to AI development as a “patchwork of 50 different regulatory regimes” — states passing their own laws instead of one federal standard. Unlike Biden’s executive orders, which lacked external legal force, Trump’s Executive Order 14179 created tangible action by establishing a Department of Justice (DOJ) AI Litigation Task Force responsible for investigating, challenging, and potentially prosecuting state-level AI legislation, “including on grounds that such laws unconstitutionally regulate interstate commerce.” This framing invokes the dormant commerce clause, which prohibits states from passing legislation that unduly burdens interstate commerce even when Congress has not acted. However, whether state AI regulation actually constitutes an undue burden on interstate commerce has not been adjudicated, and courts have historically been reluctant to strike down state consumer protection laws on dormant commerce clause grounds. The legal viability of the Task Force’s theory remains an open question. Trump’s prioritization of a unified national standard over individual state laws bears resemblance to the Airline Deregulation Act and the Clean Air Act, both of which preempted state regulation in their respective fields. The critical distinction, however, is constitutional: both of those statutes were passed by Congress, giving them preemptive force under the Supremacy Clause. Executive Order 14179 was not: an executive order attempting to achieve the same preemptive effect as a congressional statute lacks the same constitutional foundation, which is precisely why the administration must rely on the DOJ Task Force to litigate state laws one by one rather than displacing them directly.
The limits of this approach became visible in Congress itself. Before it moved to the Senate, the House of Representatives’ version of the One Big Beautiful Bill Act originally contained a provision preventing state and local governments from developing AI regulation for the next ten years. However, senators voted ninety-nine to one to remove this provision from the bill. The near-unanimity of opposition to the provision is not only a political observation, but reflects a core principle of preemption doctrine: the federal government cannot strip states of regulatory authority without first enacting a federal regulatory framework to replace it. Without a national AI law, there is nothing for the federal authority to preempt. Under the anticommandeering doctrine established in New York v. United States (1992), the federal government cannot compel states to enact or enforce federal regulatory programs. While blocking state regulation is distinct from compelling state action, the underlying Tenth Amendment principle is similar: the federal government cannot dictate how states exercise their legislative authority in areas where Congress itself has not acted. The House’s vote to block state AI regulation for ten years therefore posed Tenth Amendment concerns, as it would strip states of authority in an area where no federal standard exists to replace it. In this regulatory vacuum, AI companies are able to continue developing their technology unchecked, which is arguably the intended outcome.
The federalist struggle between the Biden and Trump administrations poses a deeper constitutional question: whether the executive branch can achieve through executive orders and litigation what would ordinarily require congressional legislation. While Biden demonstrated that executive power alone cannot create enforceable AI regulation, Trump is now demonstrating that executive power alone may not be able to prevent it either, as the ninety-nine to one Senate vote and the unresolved dormant commerce clause questions suggest. Until Congress acts, AI regulation will remain caught between an executive branch that lacks the constitutional authority to mandate it and states whose authority to fill the gap is being actively contested.
By Nikita Pande, from CULR
Section II: When Labels Vanish, How Rolling Back Model Cards Weakens Transparency for Healthcare AI
When the Trump Administration's Department of Health and Human Services (HHS) proposed eliminating model card requirements under HTI-5 on December 22, 2025, it threatened the documentation standards governing how clinical AI algorithms are trained, validated, and monitored, along with the legal frameworks that depend on them. The proposal seeks to reduce burdens and roadblocks to AI growth by implementing President Trump's Executive Order 14192, "Unleashing Prosperity Through Deregulation," which mandates that federal executive-branch agencies repeal ten existing regulations for every new one issued, slashing private expenditures required for compliance with federal regulations. The HTI-5 proposed rule seeks to eliminate thirty-four out of sixty certification requirements, affecting nearly 70 percent of existing standards set by the Assistant Secretary for Technology Policy/Office of the National Coordinator for Health Information Technology (ASTP/ONC), drastically altering compliance landscapes. By imposing the ten-to-one replacement ratio without regard for the content of each regulation, the proposed approach treats all regulations as functionally identical. Model card requirements for clinical algorithms, however, serve a critically different purpose than routine paperwork requirements, with the former being crucial to protecting patient safety. This raises the doctrinal question of whether a blanket deregulation rule is appropriate for healthcare AI.
Following the deregulation directive issued through Executive Order 14192, one of the thirty-four certification criteria the HHS's HTI-5 Proposed Rule will eliminate is the necessity of model cards for clinical decision support software. Model cards describe a machine learning model's technical construction, risks, performance metrics, and known limitations and biases. These standardized records facilitate compliance with regulations and audits, and consistent governance across models. Introduced during the Biden era, they were a product of Executive Order 14110, "Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence," which demanded model-level documentation for high-impact systems. The National Institute of Standards and Technology (NIST) AI Risk Management Framework explicitly points to model cards as a transparency tool for documenting explanatory and validation information.
The rollback of model card requirements thus takes place via two separate legal processes: a broad deregulatory directive through Executive Order 14192 and a specific deregulation proposal through HTI-5, which requires the promulgation process to follow the notice and comment procedure. Since executive orders are not regulated by the procedures stipulated in the Administrative Procedure Act, a federal law governing how U.S. administrative agencies may propose and issue regulations, the only formal forum available to challenge the rollback of the model cards would be the HTI-5 comment process. On the other hand, the fact that model cards emerged from Executive Order 14110, and are now slated for removal under Executive Order 14192, demonstrates the vulnerability of the requirement for transparency in government AI systems. If such a mandate is not codified into statute, they remain subject to reversal as executive directives. This institutional fragility proves relevant to healthcare law because the courts and hospitals cannot possibly hope to treat model cards as a stable part of the evidentiary and compliance landscape if their very existence depends entirely on shifting presidential priorities; their creation and abolition through executive-driven deregulatory rulemaking only highlights that instability.
Well-documented medical files are the foundation for American healthcare jurisprudence, with legal theories premised on the assumption that there is sufficient documentation to establish causes of medical malpractice and negligence. Courts depend on this documentation to evaluate whether patients suffered from their clinician’s poor decision making, and this need was reinforced by Gold v. Greenwich Hospital Association (2002). The court held that a hospital's failure to maintain adequate records prevented a meaningful reconstruction of the care provided, which constituted negligence on the hospital’s behalf. Similarly, Darling v. Charleston Community Memorial Hospital (1965) determined that hospitals owe their patients an independent duty of care extending beyond individual physicians to institutions themselves. Maintaining well-documented records is one way hospitals fulfill that obligation. This practice has therefore been incorporated into state statutes and medical accreditation agency policies. For example, the Florida Supreme Court’s Valcin doctrine creates a rebuttable presumption that may shift the burden of proof in medical negligence cases to the defendant in the event of missing records. If model cards are eliminated and a patient is harmed by an AI‑generated clinical recommendation, the question is whether courts would apply the adverse inference approach seen in Valcin. Because the existence of model cards would no longer be required, their absence could trigger precisely this evidentiary presumption against those parties which deployed AI in clinical care, effectively penalizing them for a documentation gap that federal deregulation itself created.
Beyond the requirements of common law doctrines, all states have explicit statutes mandating recordkeeping and record retention, linking failure to document sufficient information to a physician’s license to practice medicine. However, some states layer tort‑reform measures on these duties that constrain how patients can enforce them in malpractice litigation. For instance, Chapter 74 of the Texas Civil Practice and Remedies Code, which outlines medical liability reform laws implemented in 2003, puts a cap on noneconomic damages in any malpractice lawsuits. Section 74.351 requires the plaintiff to submit reports by one or more experts stating the methods by which the alleged malpractice caused injury within 120 days of filing the lawsuit. Eliminating model cards makes this requirement harder to satisfy: without standardized documentation on a clinical algorithm, identifying an expert who can opine on how an AI algorithm caused harm becomes significantly more difficult.
The black-box nature of machine learning models eludes such regulatory frameworks, creating a legal gray area. In practice, this means that a model’s internal decision process is practically opaque to even its creators, posing a direct conflict with informed consent doctrine. Physicians possess a legal obligation to disclose the material risks of any proposed treatment so patients can make informed decisions. If the recommendation generated by an algorithm cannot be explained by the doctor or even the model’s developers, as in a black-box scenario, the doctor will struggle to thoroughly fulfill their obligation of informed consent. Model cards enable doctors to better understand algorithms and their potential risks, so they may inform the patient accordingly. Eliminating the requirement for model cards removes the one tool that can bridge this chasm between physicians’ duty to disclose risks and the obscurity of a black-box model.
Presently, U.S. courts have not definitively ruled on the issue of whether physicians have a legal requirement to give information regarding AI-assisted decision-making to their patients. However, since informed-consent law has been extended to cover novel procedures and technologies before, it may be assumed that informed-consent law should also be extended to cover AI tools. This doctrinal gap has been increasingly relevant to the rapid integration of AI in electronic health records. Recent surveys indicate that 70 percent of American hospitals embed predictive AI tools in electronic health records, and nearly one-third use generative AI for documentation and clinical support. Malpractice claims involving AI-driven diagnosis have increased 14 percent between 2022 and 2024, yet courts struggle to determine liability because of the failure of U.S. courts to rule on the allocation of responsibility when using AI. That is, whether the physician relying upon the tool, the hospital where it was implemented, or the company behind its development should take responsibility. Under Darling's corporate-negligence doctrine, hospitals owe patients an independent duty of care to vet, monitor, and document the AI tools that they use. Model cards fulfill this obligation, providing a structured record of the mechanisms by which these systems are trained, validated, and monitored, and therefore stand as the primary means of fulfilling that institutional duty.
The HTI-5 proposal presents a key test of Trump’s deregulatory AI agenda, having the potential to reshape clinical care by compromising patient safety in favor of rapid innovation. By treating model cards as dispensable, the administration threatens to undermine patient access to information about the AI tools shaping their care, and the ability of courts to hold institutions accountable when those tools cause harm. Transparency tools like model cards are imperative to adequate clinician disclosure, which in turn stands to justify lighter ex ante regulation. Eliminating model cards would create an evidentiary gap in the adjudication of malpractice cases, weaken the ability to fulfill informed consent requirements, and dismantle the established documentation framework that healthcare law has depended on for decades.
By Sofia Grimm, from YULJ
Section III: Emerging Datascapes, The Legal Battle for Indigenous Water Rights Amid the Push for AI Data Centers
On July 23, 2025, the White House issued Executive Order 14318, “Accelerating Federal Permitting of Data Center Infrastructure,” which stated that the Trump Administration would “utilize federally owned land and resources for the expeditious and orderly development of data centers” that would be used “in service of the prosperity and security of the American people”. The construction of such data centers causes environmental, energy, and water-usage concerns on the state and federal level. Public Domain Allotments (PDAs), approved in collaboration with the Bureau of Land Management (BLM), are “land reserved out of the public domain for use by an Indian person or family”, and often surrounded by federal public land. Unlike larger reservations, PDAs are not under the jurisdiction of any tribe or Indian nation. Public Domain Allotments on their own do not reserve specific water rights for Indigenous peoples. However, as Rubin argues in the Yale Law Journal, the federally reserved water rights established in Winters v. United States (1908) should rightfully extend to PDAs, since the intent behind their creation is similar to that of reservations. As of today, this extension to PDAs is largely unrecognized and unenforced. This legal gap in extending water rights to PDAs itself becomes a major issue in light of the construction and establishment of these large AI data centers. Already, a tenth of Native Americans lack access to safe, clean drinking water due to historical legacies of colonialism and environmental destruction. Executive Order 14318 allowing AI data centers to be constructed on BLM land will exacerbate this already concerning statistic. In addressing and analyzing this legal concern, there is a clear inadequacy of federal legal frameworks to protect Indigenous water rights on PDAs in the face of Artificial Intelligence backed infrastructural threats.
Executive Order 14318 threatens an assault on the water found on federal public land. Section 9 of Executive Order 14318 aims to promote the building of AI data centers through the Department of the Interior, Department of Energy, and Department of Defense. The Order further invokes 43 U.S.C. 1701, which governs public lands retained in federal ownership and managed by the BLM as a basis for siting data center construction. The invocation of this statute provides legal authority for siting AI data centers on federal lands. In Seven County Infrastructure Coalition v. Eagle County (2025), the Supreme Court held that the National Environmental Policy Act (NEPA) only requires all federal agencies in the executive branch to review environmental impacts they have regulatory authority over. This narrows the scope of NEPA reviews and allows agencies to define “manageable lines” for these reviews, expediting agency approval and permitting processes. If NEPA reviews are limited to impacts an agency has direct regulatory authority over, then the water quality effects of a data center built on BLM land, or any agency infrastructure on a nearby PDA, could fall through the cracks between agencies and receive no comprehensive review. When Seven County Infrastructure Coalition v. Eagle County is considered alongside Executive Order 14318, new dangers to Public Domain Allotments are exposed. AI data centers could now be built on federal public land with an expedited permitting process that potentially endangers water quality, and therefore water rights, on bordering PDAs.
A PDA often has multiple beneficiaries to the land trust, typically Indigenous members of the same family. PDAs are legally distinct from reservation land and have their own specific water rights, as does federal public land. Under Winters v. United States (1908), reservations have federally reserved water rights. In Winters, the Court held that when the federal government created reservations, it implicitly reserved enough water to fulfill the reservation’s purpose, despite the treaty not explicitly mentioning water. As stated by Rubin, water rights as allocated by Winters are to be used to fulfill the “purpose of the land.” Further, PDAs should be recognized with the same federally reserved water right as the intent underlying their creation. The goal behind the creation of PDAs is similar to that of reservations, which is to support indigenous settlement and survival. Therefore, PDAs should trigger the same implied reservation of water recognized in Winters. Currently, PDAs do not have these recognized rights, suffering from what Rubin identifies as a “recognition gap” between the logic of the federal reserved rights doctrine under Winters and its application to allotments. This gap leaves PDA water resources legally unprotected against competing uses, such as AI data centers permitted under Executive Order 14318.
Tribes have the power to invoke NEPA, NHPA, and American Indian Religious Freedom Act in legal efforts to prevent the construction of these data centers. In Hualapai Indian Tribe v. Debra Haaland et al. (2024) the Hualapai tribe challenged a federal and private lithium drilling project near a sacred site on federal land, the U.S. District Court for the District of Arizona granted a preliminary injunction to their case. This court found that the tribe was likely to succeed on its claims under NEPA and the NHPA. While the Hualapai preliminary injunction suggests that sacred tribal sites and resources can halt federal projects under NEPA and NHPA, Lyng v. Northwest Indian Cemetery Protective Association (1988) set the opposing precedent: the government can build infrastructure that “benefits the public good” on sacred tribal land. Under Lyng, one could argue that data centers should be considered a public good in the face of Indigenous wellbeing and religious practices, given the extent of AI’s integration into American society.
Indigenous water rights are under threat as the construction of AI centers mandated by Executive Order 14318 may be justified as infrastructure that benefits the common good. Many of these definitions of the public good within takings jurisprudence have historically favored economic development and productivity over competing land uses. In Kelo v. New London (2005), the Supreme Court broadly interpreted “public use” within the meaning of the Takings Clause. The Supreme Court ruled that a city seizing private property for a private economic development project constituted justification through the definition of “public use” under the Fifth Amendment because it served a public purpose, such as increasing tax revenue and employment. While PDAs are held in federal trust rather than as private property, the broad definition of “public use” poses the concern that similar reasoning could be applied to justify federal infrastructure on or near PDA land. Namely, under Kelo, the seizure of PDAs for the development of data centers could be justified through protections for “public use” or “public good” under the Fifth Amendment as these centers will likely be hubs for employment and increasing tax revenue.
The compounding tragedy that state and legal structures in the United States have imposed upon Indigenous peoples is clear though the lack of legal protections of their land and their people against the emerging artificial intelligence datascape. Considering Winters in conjunction with the Executive Order 14318, it is evident that there is a gap in legal frameworks to Indigenous water rights in light of the shifting technological and natural landscape. The Winters recognition gap must be closed for PDAs, and the narrowing NEPA framework under Seven County must be addressed before Executive Order 14318 does irreversible damage. Closing this gap requires a judicial decision explicitly extending Winters reserved water rights to PDAs, paired with legislative action to ensure that NEPA review processes account for the cumulative water impacts of federal infrastructure projects on adjacent indigenous lands.
By Jane Bryant, from CULR
Section IV: Human Out of the Loop
Picture a weapon that decides, on its own, who to target. No human makes the decision and no commander reviews the choices; instead, an algorithm processes inputs and outputs that lead to loss of life. Companies such as Palantir and Anduril are developing technologies capable of selecting and engaging targets without human oversight at the moment of execution; the legal implications demand scrutiny. Fully autonomous weapons systems, by their nature, cannot satisfy the fundamental requirements of distinction, proportionality, and precaution under International Humanitarian Law (IHL), making their use a violation of the Geneva Conventions and Additional Protocols irrespective of advances in algorithmic design. This essay analyzes how autonomous weapons fail each of these principles in turn, before addressing the accountability gap their use creates under international criminal law.
IHL governs conduct in armed conflict and is primarily established in the Geneva Conventions and their Additional Protocols. It requires three principles to be respected during armed conflict: distinction, proportionality, and precaution. Distinction, codified in Article 48 of Additional Protocol I, requires that parties to a conflict “at all times distinguish between the civilian population and combatants, between civilian objects and military objectives, and accordingly direct their operations only against military objectives.” Proportionality, established in Article 51(5)(b), prohibits attacks that may cause “incidental loss of civilian life, injury to civilians, damage to civilian objects, or a combination thereof” when such harm would be excessive in relation to the concrete and direct military advantage anticipated. An attack is unlawful if civilian harm is too high relative to the military gain achieved. Proportionality is ultimately a moral judgment requiring the weighing of a multitude of contextual factors, such as the presence of a child standing next to armed men, that no algorithm can reliably perform. Precaution, established in Article 57, requires those planning and carrying out attacks to “do everything feasible to verify that the objectives to be attacked are neither civilians nor civilian objects” and to “take all feasible precautions in the choice of means and methods of attack with a view to avoiding, and in any event minimizing, incidental loss of civilian life.” For a human commander, “everything feasible” might include pausing an attack when new intelligence emerges, redirecting a strike when civilians enter the target area, or canceling an operation when information becomes ambiguous. An autonomous system executing pre-programmed instructions cannot pause, redirect, or cancel based on real-time contextual changes. The precaution standard assumes a decision-maker capable of responding to the unexpected. An autonomous weapon, by definition, can only respond to what it was programmed to anticipate.
These principles are already being tested by the integration of artificial intelligence (AI) into military operations. Semi-autonomous weapons, often referred to as “human in the loop” systems, are deployed in international conflicts, including the war between Ukraine and Russia. Palantir has developed reconnaissance drones used in Ukraine that rely on AI software to identify personnel, tanks, and other military targets while providing targeting options to commanders. These commanders remain informed and make the final decision on whether to engage. These systems satisfy IHL’s requirement for human judgment because a human operator evaluates each target and makes the final decision to use lethal force. However, defense companies are now developing fully autonomous, “human out of the loop” weapons that can execute a target without requiring human approval. Anduril’s development of AI-piloted aircraft for the U.S. military represents this next step: removing human oversight entirely. This creates a direct conflict with Article 48’s requirement that parties distinguish between civilians and combatants at all times. The question is not whether AI can assist targeting, but whether it can replace the human judgment IHL demands at the moment of lethal action.
The U.S. Department of Defense has already created a framework that permits this replacement. Directive 3000.09 defines Lethal Autonomous Weapons Systems (LAWS) as “weapon system[s] that, once activated, can select and engage targets without further intervention by a human operator.” The directive permits such systems under conditions requiring “appropriate levels of human judgment,” but this standard is too vaguely defined to satisfy IHL. Under Article 48, distinction requires a human assessment of each specific target at the point of engagement. The directive’s standard could permit human judgment only at the programming or activation stage, allowing the system to autonomously select and engage individual targets thereafter. A blanket authorization to engage categories of targets is not the same as a human evaluating whether a specific individual is a combatant. IHL requires the latter, and the directive does not guarantee it. Furthermore, a domestic policy framework cannot lower the threshold established by treaty obligations; the United States remains bound by the distinction requirement of Additional Protocol I regardless of how Directive 3000.09 defines “appropriate levels of human judgment.”
The technical limitations of autonomous systems compound these legal failures. A 2024 paper by the Stockholm International Peace Research Institute (SIPRI), titled “Bias in Military Artificial Intelligence,” found that these AI systems are vulnerable to systemic biases that come through context, development, and data. If an algorithm's training data causes it to disproportionately flag individuals of a certain demographic as combatants, the system cannot reliably satisfy the distinction Article 48 requires. Systemic bias then becomes a mechanism through which the legal requirement of distinction is violated, producing targeting outcomes that disproportionately harm civilian populations along the lines of race, social context, or gender.
These technical limitations are equally fatal to the proportionality requirement. A Human Rights Watch Report from 2025 states, “Algorithms cannot adequately interpret human behavior, intentions, or rapidly changing circumstances” and cannot make proper proportionality judgments. The inability to interpret human behavior matters legally because proportionality under Article 51(5)(b) requires weighing civilian harm against military advantage in a real-time context. A combatant surrendering, a civilian picking up a weapon in self-defense, a child wandering into a strike zone: these are contextual developments that proportionality demands decision-makers account for. An algorithm processing pre-programmed parameters cannot adapt to circumstances it was not trained to recognize, meaning it cannot fulfill the weighing that the legal standard requires.
Beyond the failure to satisfy distinction, proportionality, and precaution, autonomous weapons create a fundamental gap in legal accountability. Under IHL, individual commanders bear personal responsibility for ensuring that attacks comply with these three principles. Under Article 30 of the Rome Statute, individual criminal liability requires that a person acted with “intent and knowledge.” A commander who orders a strike intends a specific outcome and can be held accountable if that strike violates IHL. When an autonomous weapon system makes the targeting decision, no human exercises intent at the moment of lethal action. If no human chose the target, no individual satisfies the mens rea threshold required for criminal liability. States or manufacturers may face responsibility after the fact, but IHL’s framework of individual accountability depends on a human decision-maker who intended the outcome. Autonomous weapons eliminate that decision-maker, creating a gap that current international criminal law cannot close. Supporters of autonomous weapons argue that the use of algorithms could reduce civilian casualties by eliminating any human error potentially involved in targeting decisions. However, this argument misunderstands what is required under IHL: not simply accurate targeting, but also human moral judgment and legal accountability at the moment of lethal force.
When an autonomous weapon system makes critical targeting decisions, there are no human beings applying the principles of distinction, proportionality, and precaution, thereby undermining the core safeguards of IHL. Currently, the Geneva Conventions and their Additional Protocols have not been amended to address autonomous systems. They should be. An additional protocol should establish that any weapon system making targeting decisions must have a human operator who exercises judgment over each individual engagement, ensuring that the principles of distinction, proportionality, and precaution are applied by a person who can be held legally accountable under the Rome Statute. Without such an amendment, the gap between what IHL requires and what autonomous weapons can deliver will only widen as the technology advances.
By Milad Davaran, from YULJ
This piece was edited by Qizhen (Kiara) Ba and Jasmine Rocha from CULR and Drea Cabral and Ruby Knoebel from YULJ.