The previous chapter examined data fusion as a structural problem in immigration enforcement: the legal consequences that arise when fragmented records from multiple sources are combined into a single operational picture. This chapter examines a different but related problem — what happens when that data is used not to document the past but to predict the future.
Predictive risk scoring in immigration is not hypothetical. It is operational. DHS’s AI use-case inventory lists Hurricane Score — identifier DHS-2408 — as a machine-learning-based risk indicator used within ICE’s Alternatives to Detention program, specifically the Intensive Supervision Appearance Program known as ISAP, to assess the likelihood that an individual under supervision may abscond. The system produces a score from 1 to 5. Higher scores indicate greater model-estimated risk. Officers are instructed to consider the score alongside other case information when deciding whether to adjust supervision conditions.
The legal question this chapter addresses is not whether the system should exist. It is what legal framework governs it, what rights it implicates, and what tools a lawyer has to challenge a supervision decision that the score may have shaped.
I. The technical structure of Hurricane Score and its legal classification
DHS AI Use Case Inventory, DHS-2408
The DHS inventory entry for Hurricane Score is detailed enough to support a legal analysis of the system’s character and governance obligations.
The model is a quasi-binomial, binary classification machine-learning system trained on historical ATD-ISAP case data — specifically, data from inactive cases — to identify patterns associated with prior absconding behavior. Its inputs include case-specific variables: the number of compliance violations recorded, the length of time the individual has been enrolled in the program, and whether the individual holds a valid travel document. The output is a periodic score, generated at case review intervals, used to inform decisions about the appropriate level of case management and technological monitoring assigned to each enrollee.
DHS designates Hurricane Score as a rights-impacting AI system — meaning the agency itself acknowledges that the model’s output can affect the legal rights or interests of individuals. The inventory also identifies known risk factors: the model may incorrectly assign elevated scores to individuals who are not at genuine risk of absconding, and the training data may be insufficiently representative of the current supervised population, leading to degraded performance for some subgroups.
That self-description is the starting point for any legal challenge. The government has acknowledged in its own official documentation that the system produces risk-impacting outputs, that false positives are a known failure mode, and that training data quality is a recognized concern.
II. The applicable US legal framework — Fifth Amendment and Mathews
U.S. Constitution, Amendment V Mathews v. Eldridge, 424 U.S. 319 (1976)
The central US legal framework for challenging a predictive immigration score is procedural due process under the Fifth Amendment, analyzed through the Mathews v. Eldridge balancing test established in Chapter 12.
Mathews requires a court to weigh three factors: the private interest at stake, the risk of erroneous deprivation under existing procedures and the value of additional safeguards, and the government’s interest. Applied to Hurricane Score, the analysis proceeds as follows.
The private interest is significant. ISAP supervision can include electronic monitoring, GPS ankle bracelets, telephonic check-ins, and intensive case management requirements. An elevated score can result in movement toward detention or in heightened surveillance conditions that substantially constrain liberty and interfere with employment, family life, and legal proceedings. That is not a minor administrative consequence.
The risk of erroneous deprivation is structurally elevated. The person subject to the score typically receives no notice that the score exists, no disclosure of the specific inputs as applied to their case, and no formal mechanism to contest the score’s accuracy or reliability before supervision conditions are adjusted. Where the inputs are inaccurate — stale compliance records, unresolved violation entries, incorrect travel-document status — the score may be wrong in ways that the affected person cannot identify or challenge. The DHS inventory’s own acknowledgment of false-positive risk confirms that the risk of error is not speculative.
The government’s interest in efficient supervision management is legitimate. But it does not require that supervision decisions be made on the basis of opaque inputs the person cannot examine. Notice, access to the relevant inputs, and an opportunity to correct demonstrable factual errors impose modest administrative costs relative to the liberty interests at stake.
Under Mathews, the combination of significant private interest and elevated structural risk of error supports the argument that due process requires, at minimum, that the individual be informed that a risk score exists, that they be provided the inputs used to generate it, and that they have a meaningful opportunity to challenge factual inaccuracies in those inputs before supervision conditions are materially changed.
III. Predictive scoring versus recordkeeping — a structural distinction
A standard immigration file contains verifiable facts: dates of entry, past encounters, application history, court records, compliance entries. Those facts can be reviewed, contested, and corrected through standard administrative and judicial processes.
A predictive score does something different. It applies a statistical model to those facts and produces a probabilistic judgment about future behavior. That transformation raises a legal problem that factual recordkeeping does not. The person subject to a prediction is not defending against something they did. They are defending against a model’s inference about what they might do — an inference generated by design choices about which variables correlate with non-compliance, how those variables are weighted, and whether the training population adequately represents them as an individual.
If the government cannot explain those design choices in a legally meaningful way — the selection of variables, their weighting, the validation of the model against the current supervised population — then the affected person is not contesting evidence. They are contesting a black box. And as the framework we established in Chapters 2 and 13 confirms, a black box cannot serve as an adequate basis for decisions that affect liberty without the procedural protections that allow its reliability to be tested.
IV. The EU AI Act framework — what pre-deployment governance requires
EU AI Act, Regulation (EU) 2024/1689, Annex III, Articles 10, 13, 14, and 27
Under the EU AI Act, a system with the characteristics of Hurricane Score would fall within the high-risk classification under Annex III as an AI system used in migration management for risk assessment purposes. That classification triggers a set of pre-deployment and ongoing compliance obligations that illustrate the governance gap between the EU framework and current US practice.
Article 10 requires that training, validation, and testing datasets be subject to appropriate data governance practices, including examination of possible biases and assessment of the data’s representativeness for the intended purpose. The DHS inventory’s acknowledgment that Hurricane Score was trained on historical inactive-case data and that training data quality is a known risk factor maps directly onto the concerns Article 10 is designed to address. Under the EU framework, that risk would need to be documented and mitigated before deployment, not discovered after the system has been operating at scale.
Article 13 requires high-risk systems to be sufficiently transparent that deployers can interpret the outputs and use them appropriately. A score of 1 to 5 with three stated aggregate input factors and no individual-level explanation of how those factors were weighted in a specific case does not obviously satisfy that requirement. The officer receiving the score is being told what the model concluded, not why it concluded it for this individual.
Article 14 requires effective human oversight — meaning a human decision-maker who is capable of understanding the system’s limitations and who can choose to disregard or override the output where appropriate. A supervision officer who has no information about the model’s known failure modes, no visibility into the specific inputs for the case before them, and no training in how to evaluate machine-learning outputs is not exercising effective oversight in the Article 14 sense.
Article 27 requires that public-sector deployers of high-risk systems carry out a fundamental-rights impact assessment before deployment, identifying the specific rights of individuals who may be affected and the measures in place to address those risks. The DHS inventory’s designation of Hurricane Score as rights-impacting suggests that a rights assessment of some kind was conducted. Whether that assessment meets the substantive standard Article 27 contemplates — a genuine identification of risks and measures, not a pro forma declaration — is a question the inventory does not answer.
The comparison is not intended to suggest that US immigration agencies are bound by EU law. It is intended to show what an adequate pre-deployment governance framework demands, so that lawyers can identify the gaps in the current US framework and frame those gaps as legal arguments.
V. The Privacy Act — scope, limitations, and practical strategy
Privacy Act of 1974, 5 U.S.C. § 552a
The Privacy Act provides a right of access and amendment for records held by federal agencies in systems of records. Section 552a(a)(2) defines the term “individual” for purposes of those rights as a citizen of the United States or an alien lawfully admitted for permanent residence. Many noncitizens in immigration proceedings fall outside that statutory definition.
That limitation is real and must be stated clearly. It does not mean the Privacy Act is irrelevant in immigration AI cases. It means the lawyer must identify precisely what rights the client holds based on their specific immigration status, and must combine Privacy Act arguments with FOIA requests, administrative access arguments, and litigation-based discovery where available.
For clients who do qualify — lawful permanent residents and US citizens subject to immigration proceedings — the Privacy Act provides access to records in the relevant systems of records and the right to request amendment of records that are not accurate, relevant, timely, or complete. ICE’s system of records DHS/ICE-009 (External Investigations) is subject to law enforcement exemptions under 5 U.S.C. § 552a(j)(2) and (k)(2), which means access requests may be denied for records tied to active or closed enforcement matters. Even within that exemption structure, however, narrowly framed requests focused on demonstrably incorrect factual data — a wrong date, a misattributed violation, an incorrect travel-document entry — are harder for the agency to deny than broad requests for the entire case record.
The practical rectification strategy is therefore targeted. The argument is not that the algorithm is unreliable in the abstract. It is that specific, demonstrable factual errors in the model’s inputs have produced an unreliable output in this case. Violation counts, compliance history, and travel-document status are data points that can be wrong — and if they are wrong, the score built on them does not support the supervision conditions imposed on the basis of it.
VI. The legal principle: a score is not a finding
The chapter’s practical conclusion is also its legal principle.
A predictive risk score is not a judicial finding, a statutory standard, or sworn evidence. It is a probabilistic inference generated by a statistical model applied to historical patterns. Its authority depends entirely on the reliability of its inputs and the validity of the model’s design for the individual before the decision-maker. Where either of those conditions is unverified or demonstrably deficient, the score cannot serve as an adequate substitute for individualized judgment.
DHS’s own documentation of Hurricane Score acknowledges that the system is rights-impacting, that false positives are a known risk, and that the score is advisory rather than determinative. Each of those acknowledgments is also a legal argument. A rights-impacting system that produces known false positives and whose inputs cannot be examined by the affected person does not satisfy the procedural requirements that Mathews demands for decisions with significant liberty consequences. And a score that is described as advisory cannot substitute for the individualized human judgment that both due process and the government’s own policy require.
Next: Chapter 19 — AI in criminal pretrial: bail algorithms, risk scores, and the automation of liberty.
Leave a Reply