There is something almost paradoxical about the American approach to AI governance. The United States has produced most of the technology that is now transforming criminal justice, immigration enforcement, and public surveillance worldwide. And yet it remains, in 2026, without a comprehensive federal statute governing that technology’s use. What fills the gap — imperfectly, partially, but powerfully — is a document drafted in 1787.
The Constitution was written by men who had never imagined machine learning. But they had thought deeply about a different problem: how to prevent the state from using information and institutional power to override individual liberty without accountability. That problem turns out to be exactly the one that algorithmic governance poses. The tools are new. The threat is familiar.
This is where the American approach diverges fundamentally from the European one. The EU Charter of Fundamental Rights establishes constitutional limits on AI before systems are deployed, operating within a regulatory framework that imposes affirmative obligations on providers and deployers. The American constitutional framework operates differently — it is reactive rather than preventive, triggered by state action after deployment, tested through litigation rather than compliance. It is fragmented where the European framework is unified. But it is also, in the right hands, deeply powerful.
The Fourth Amendment — algorithmic surveillance and the reasonable expectation of privacy
U.S. Constitution, Amendment IV (1791)
The Fourth Amendment protects individuals against unreasonable searches and seizures by the government. For most of American legal history, its application was relatively straightforward: a physical space, a tangible object, a targeted individual. The doctrinal framework that still governs its application today was established in Katz v. United States, 389 U.S. 347 (1967), where the Supreme Court held that the Fourth Amendment protects people rather than places, establishing the concept of a reasonable expectation of privacy. What the government cannot access without a warrant is not defined by physical boundaries but by the expectations that society recognizes as reasonable.
That framework was adequate for a world where surveillance was expensive, labor-intensive, and necessarily selective. It is under serious strain in a world where algorithmic systems can aggregate license plate reader databases, public and private camera feeds, biometric identifiers, geolocation records, and social media activity — and analyze them simultaneously, at scale, in real time.
The Supreme Court recognized the early edges of this problem in Carpenter v. United States, 585 U.S. 296 (2018), where it held that accessing 127 days of historical cell-site location data constitutes a search under the Fourth Amendment. Chief Justice Roberts, writing for a 5-4 majority, rejected the argument that information shared with a third party automatically loses constitutional protection. The location data, he wrote, created a “detailed chronicle of a person’s physical presence compiled every day, every moment over years.” The comprehensive nature of the surveillance was what brought it within the Fourth Amendment — not any single piece of data, but the mosaic they formed together.
Carpenter did not address artificial intelligence directly. But its reasoning is the foundation for the most significant Fourth Amendment argument available against algorithmic surveillance. When AI systems aggregate large volumes of data — each piece individually innocuous — to generate detailed behavioral profiles, they may recreate the same kind of comprehensive monitoring that Roberts found constitutionally problematic. The “mosaic theory” of the Fourth Amendment, which several justices have discussed in various contexts, holds that even individually public or non-private data points can, when aggregated, reveal intimate details of a person’s life that carry a reasonable expectation of privacy. Applied to modern algorithmic surveillance systems, that theory has significant implications for facial recognition networks, social media monitoring tools, and the integrated data aggregation platforms now deployed in immigration enforcement.
Facial recognition illustrates the tension most clearly. Observing a person in a public space has traditionally not been considered a search. But automated identification and real-time tracking of individuals across thousands of cameras — as happens in the systems we will examine in Chapter 27 on Clearview AI — raises a different question. It is not the single observation. It is the permanent, scalable, automated record of everywhere a person has ever been.
The Fifth Amendment — due process and algorithmic decision-making
U.S. Constitution, Amendment V (1791)
The Fifth Amendment prohibits the federal government from depriving any person of life, liberty, or property without due process of law. The guarantee has two dimensions — procedural and substantive — and both become directly relevant when AI systems influence government decisions affecting individual liberty.
Procedural due process requires notice of the government action affecting the person and a meaningful opportunity to be heard before an impartial decision-maker. The word “meaningful” is where algorithmic systems create problems. Notice that an algorithm contributed to a decision is not the same as notice of how the algorithm reached its output or what data it used. A hearing that produces no mechanism for challenging the algorithmic assessment is a hearing in form but not in substance.
State v. Loomis, 881 N.W.2d 749 (Wis. 2016), which we examined in depth in Chapter 2, is the most documented illustration of this tension in the criminal context. The Wisconsin Supreme Court permitted the use of the COMPAS risk assessment in sentencing while requiring that judges receive warnings about its proprietary nature, its potential for bias, and its reliance on group-based statistics. The decision acknowledged that the defendant could not fully examine the tool that had contributed to his six-year sentence — and nonetheless found that due process was satisfied. The unresolved question it left behind is the one that makes Fifth Amendment litigation in this area so consequential: at what point does the opacity of an algorithmic system become incompatible with the constitutional right to a meaningful hearing?
In immigration proceedings, the Fifth Amendment plays a distinct and important role. The Supreme Court has long held that non-citizens in removal proceedings are entitled to due process — that the Amendment’s protection applies to “persons,” not only citizens. When algorithmic tools influence detention or deportation decisions without providing the individual any mechanism to examine the underlying data or challenge the methodology, serious Fifth Amendment questions arise. The ICE detention tools we examined in Chapter 3 — systems configured to recommend detention at rates that effectively replaced human judgment — are the clearest example. A process that looks like a hearing but whose outcome is predetermined by an unexaminable algorithm is not due process.
The Sixth Amendment — confronting algorithmic evidence
U.S. Constitution, Amendment VI (1791)
The Sixth Amendment guarantees the accused in criminal prosecutions the right to confront witnesses against them. This Confrontation Clause is the constitutional foundation of the adversarial system: you have the right to face the evidence against you, to question its reliability, to expose its weaknesses. It is the mechanism by which the legal system tests the credibility of the state’s case.
Algorithmic evidence — facial recognition matches, predictive policing alerts, risk assessment scores, forensic software outputs — creates a confrontation problem that courts have not yet resolved. The logic is straightforward: if the algorithm’s output is used against a defendant, the defendant must be able to challenge it. Challenging it requires access to information about how it works — the training data, the model architecture, the validation methodology, the error rates for the relevant demographic. Most of that information is proprietary. Companies routinely assert trade secret protection to prevent disclosure, arguing that their source code and model design are confidential business information.
The result is a direct collision between intellectual property law and the Sixth Amendment. Courts have taken inconsistent positions. Some have permitted limited review of proprietary systems under protective orders. Others have accepted algorithmic outputs as evidence without requiring any meaningful disclosure of the underlying methodology — a result that would be unthinkable with human expert testimony. You cannot present an expert witness whose methodology the defense cannot examine. The question of why an algorithm should receive different treatment is one that constitutional litigation is slowly beginning to force courts to answer.
The problem is sharpest with forensic software. Several defendants have been convicted based in part on outputs from probabilistic genotyping tools — statistical DNA analysis software — whose source code courts have refused to order disclosed on trade secrecy grounds. The same pattern appears with facial recognition evidence. We will examine this in detail in Chapter 29 on AI evidence in court, where the Daubert standard and its application to algorithmic outputs is the central subject. The Sixth Amendment argument is the constitutional foundation of that analysis.
The Fourteenth Amendment — equal protection and algorithmic bias
U.S. Constitution, Amendment XIV (1868)
The Fourteenth Amendment guarantees that no state shall deny any person the equal protection of the laws. It is the constitutional foundation for challenges to discriminatory government practices — the provision that prohibits the state from treating people differently on the basis of race, ethnicity, or other protected characteristics.
Algorithmic systems raise difficult questions under this framework precisely because the bias we documented in Chapter 3 — the COMPAS disparities, the facial recognition error rates, the discriminatory effects of immigration risk scoring — does not typically arise from explicit discriminatory design. It arises from training data that encodes historical patterns of discrimination, from proxy variables that correlate with protected characteristics, from feedback loops that amplify existing disparities. The funhouse mirror reflects the shape it was given. The discrimination is structural, not intentional.
And there is the constitutional problem. In Washington v. Davis, 426 U.S. 229 (1976), the Supreme Court held that the Equal Protection Clause requires proof of discriminatory intent, not merely discriminatory impact. A practice that produces racially disparate outcomes is not unconstitutional solely because of those outcomes. The plaintiff must show that the disparity reflects purposeful discrimination.
This intent requirement makes constitutional equal protection challenges to algorithmic systems extremely difficult. A system trained on historical criminal justice data, producing racially disparate risk scores, with no evidence that anyone designed it to discriminate, may not satisfy the Washington v. Davis standard regardless of how severe its disparate impact is. That gap — between the documented harm and the available constitutional remedy — is one of the most consequential unresolved problems in American AI governance law. Plaintiffs have argued that continued government reliance on systems known to produce biased outcomes, in the face of documented evidence of those outcomes, amounts to something closer to deliberate indifference. Some courts have been receptive to that framing. The doctrine remains unsettled.
It is worth noting that statutory claims under Title VI or the Civil Rights Act may offer a more viable path in some contexts, because those statutes do permit disparate impact claims without requiring proof of intent. The constitutional argument and the statutory argument are not alternatives — they are complementary, and the strongest litigation strategies deploy both.
Due process as the closest U.S. equivalent to rectification rights
The United States does not have a general legal right to correct inaccurate personal data held by government systems. There is no GDPR Article 16 equivalent in American constitutional law. But due process under the Fifth and Fourteenth Amendments can provide functionally similar protection in specific contexts — and for lawyers representing clients whose lives have been affected by algorithmic errors, understanding those mechanisms is operationally essential.
In criminal proceedings, discovery procedures, motions to suppress, and Brady obligations — which require the government to disclose exculpatory material, including information that undermines the reliability of evidence — can be used to challenge erroneous data in algorithmic systems. In immigration proceedings, the Freedom of Information Act and the Privacy Act of 1974, 5 U.S.C. § 552a, allow individuals to request records held by federal agencies and seek corrections of inaccurate information. These mechanisms are narrower, more procedurally demanding, and less structurally coherent than the GDPR’s correction rights. But they exist, they have been used successfully, and we will examine how to deploy them in detail in Chapters 22 and 31.
Building a constitutional litigation strategy
The Fourth, Fifth, Sixth, and Fourteenth Amendments provide a powerful but fragmented constitutional arsenal for challenging government use of AI. A comprehensive strategy rarely relies on a single amendment. The strongest cases combine multiple constitutional arguments — Fourth Amendment challenges to the surveillance that generated the data, Fifth Amendment arguments demanding transparency and meaningful participation, Sixth Amendment claims targeting the admissibility of algorithmic evidence, Fourteenth Amendment equal protection challenges to discriminatory outcomes — with parallel statutory claims and, where available, state law remedies.
The strategic framing that runs through all of these arguments is the same: artificial intelligence is not neutral technology. When deployed by the government to make or influence decisions affecting people’s liberty, it is state action. And state action is subject to constitutional scrutiny regardless of the sophistication of the tool the state is using.
That reframing — from technical system to state actor — is where constitutional AI litigation begins. The amendments provide the framework. The facts of the specific system, the specific deployment, and the specific harm provide the substance. The chapters that follow build the evidence.
Leave a Reply