Resources

For most of modern immigration history, adjudication in the United States has been understood as an inherently human process. Immigration officers reviewed applications, evaluated testimony, weighed evidence, and exercised discretion in decisions that could determine whether an individual remained in the United States or returned to danger abroad. Even when outcomes were contested, the system rested on an assumption: that human judgment, however imperfect, remained at the center of decision-making.

There were systems and checks in place to account for or correct errors. Immigration processes are now changing. We no longer know how decisions are being made and whether those same systems and checks will continue to guarantee fundamental fairness and due process.

The Department of Homeland Security (DHS) has publicly confirmed, through its Artificial Intelligence Use Case Inventory (Artificial Intelligence Use Case Inventory | Homeland Security, that artificial intelligence and machine learning tools are now integrated into numerous operational functions within U.S. Citizenship and Immigration Services (USCIS). These tools are described as mechanisms to improve efficiency, reduce backlogs, and assist officers in managing an unprecedented volume of applications. DHS emphasizes that human adjudicators retain decision-making authority and that AI systems do not independently grant or deny immigration benefits.

Yet the introduction of algorithmic tools into administrative processes raises a different and more complex question. Even where humans retain final authority, how do automated systems shape the environment in which those decisions are made? What influence do automated classifications, prioritization tools, or similarity analyses have on officer perception? And what safeguards exist to ensure that efficiency-driven automation does not unintentionally alter outcomes in a system where individual decisions carry extraordinary consequences?

Attorney Observations: Unexplained Outcomes and the Invisible Hand of AI

Immigration attorneys across the country are increasingly reporting cases that seem inexplicable, inconsistent, or unusually harsh. Various social media reports highlight some of these examples. One practitioner shared an instance in which an approval notice was mailed to a third party’s address entirely unrelated to the applicant, prompting speculation that an automated document-handling system may have misassigned records. Other attorneys report RFEs that cite “similarity concerns” or “inconsistencies” without clear explanation, raising questions about whether automated text analytics or internal flagging mechanisms are influencing officer attention or priorities.

The American Immigration Lawyers Association (AILA) has documented patterns of inconsistent adjudication where reasoning in the record does not always align with submitted evidence. Although these anomalies are not direct proof of AI errors, they underscore a broader challenge: applicants and their attorneys often cannot see how decisions are made, nor whether emerging technologies are shaping outcomes behind the scenes.

The concern is not speculation about secret decision-making, but about a system where AI-assisted processes may silently influence which cases receive more scrutiny, which narratives are flagged, and which applications are prioritized, all while remaining invisible to the public and courts. In a high-stakes environment where a single denial can upend a life, even small, opaque shifts in workflow can carry profound consequences.

Understanding the DHS AI Inventory: What “29 Use Cases” Actually Means

The DHS Artificial Intelligence Use Case Inventory lists approximately twenty-nine USCIS-related AI use cases as of its February 2025 update. For readers unfamiliar with government technology inventories, this number is easily misunderstood.

A “use case” does not refer to the number of times artificial intelligence has been used, nor does it correspond to individual immigration applications. Instead, each use case represents a distinct operational system or context in which AI or machine learning is deployed, tested, or evaluated. A single use case may operate across thousands or even millions of cases, often automatically and without direct user interaction.

For example, a document-classification system may analyze every piece of evidence uploaded into USCIS electronic filing systems. An identity-deduplication model may run automatically whenever biometric or biographical information is processed. In this sense, the inventory describes categories of AI deployment rather than the scale of usage.

The inventory categorizes systems as deployed, pre-deployment, or inactive, and identifies certain systems as “rights-impacting,” meaning they are subject to additional internal risk-management requirements under federal AI governance guidance. DHS states that these systems are designed to support processing rather than replace human decision-making.

The distinction between assistance and automation is central to understanding both the promise and the risks of AI in immigration adjudication.

AI in Adjudicative Workflows: How Influence Differs from Decision-Making

DHS materials repeatedly emphasize that AI tools used by USCIS do not make benefit determinations. Public descriptions indicate that current deployments are primarily used for:

• Organizing and classifying evidence
• Identifying duplicate identity records
• Assisting fraud detection workflows
• Biometric identity verification
• Translation experimentation
• Internal analytics or training support

In other words, DHS says that AI is embedded not in final decision-making, but in adjudicative workflow or the sequence of steps through which an application moves before a decision is made. These functions may appear administrative rather than adjudicative. However, administrative tools influencing workflow can indirectly shape outcomes even when final decisions remain human.

Artificial intelligence can affect which files are reviewed first, which issues are highlighted, how evidence is grouped, or which elements of an application receive greater attention. In cognitive science, this is often described as shaping the “decision environment.” The order in which information is presented and the signals associated with that information can influence human judgment.

The question, therefore, is not whether AI replaces adjudicators. It is how automated systems influence the informational context within which adjudicators exercise judgment.

Human Oversight: Policy Assurances and Practical Unknowns

DHS policy requires human oversight for rights-impacting systems, and agency materials consistently state that AI outputs are advisory. However, publicly available documentation provides limited detail regarding how oversight operates in practice.

For example, DHS disclosures do not explain:

• Whether officers see neutral organizational outputs or affirmative risk indicators
• Whether automated analyses include confidence scores or recommendations
• How frequently officers override automated flags
• Whether officers must conduct independent document review of AI-identified issues
• Or whether audit data exists showing how long officers review cases after automated processing

Modern case management systems typically record timestamps and user activity for security and auditing purposes, meaning such data may exist internally. However, DHS has not publicly released metrics that would allow external evaluation of how human oversight functions operationally.

This absence of detail does not establish wrongdoing. It does, however, make meaningful public assessment of safeguards difficult. In high-volume administrative systems such as immigration case processing, how the AI generated information is presented, not individual analysis of the human adjudicator, can often determine whether automated assistance enhances or undermines decision quality.

Backlogs, Scale, and the Institutional Incentive to Automate

The expansion of AI within USCIS cannot be understood apart from the scale of the modern immigration system. Millions of immigration benefit applications remain pending across multiple categories, and affirmative asylum processing alone involves hundreds of thousands of cases. Immigration courts face similar historic backlogs and a large number of new adjudicators.

As of Q3 FY2025 (April–June 2025), USCIS reported a pending application number of over 11 million cases across all application types.

Under these conditions, efficiency becomes an institutional necessity. Tools capable of organizing large volumes of evidence, identifying duplicate records, or streamlining file review offer clear operational benefits.

What remains unknown is whether AI-assisted workflows alter adjudicative outcomes. DHS has not published outcome-based performance metrics comparing AI-assisted processing to traditional processing, nor has it released data evaluating whether automated assistance changes approval rates, denial rates, or error rates. Without such information, it is difficult to assess the tradeoffs between speed and accuracy.

The absence of outcome data does not demonstrate harm. It does, however, limit the ability of the public and policymakers to evaluate the real-world effects of automation.

Lessons from Other Government AI Systems

Concerns about automated influence in administrative decision-making are not unique to immigration. Over the past decade, multiple government systems have incorporated algorithmic tools with mixed results.

Criminal justice risk assessment algorithms, most notably the COMPAS system, prompted litigation and public debate after defendants challenged the opacity of scoring methodologies used in sentencing contexts. Automated unemployment eligibility systems in several states later generated large numbers of erroneous fraud determinations, leading to widespread appeals and settlements.[xii]Healthcare authorization algorithms used in Medicare Advantage programs have likewise faced litigation alleging overreliance on automated determinations.

These examples do not establish similar problems within USCIS. They do, however, demonstrate a recurring pattern: automated systems often operate without controversy until their effects become visible through litigation or investigative reporting. Governance questions frequently emerge only after implementation and patterns emerge. 

Immigration adjudication, given its scale and consequences, may now be encountering similar structural questions.

Bias Risk and the Limits of Public Evaluation

There is no public evidence that USCIS AI systems use nationality, surname, or ethnicity as predictive variables. Nevertheless, extensive research across industries shows that machine learning systems trained on historical data can reproduce existing patterns unless explicitly corrected.

The concern in immigration contexts is therefore not proven bias but the possibility that historical enforcement, investigative patterns, or ongoing AI learning could unintentionally shape automated outputs. DHS states that bias testing and mitigation procedures exist for certain systems, particularly those involving biometrics, but performance metrics and testing methodologies are not publicly disclosed.

Without access to testing results or independent audits, external observers cannot meaningfully evaluate how bias risks are identified or mitigated over time.

Translation, Language, and Traceability Concerns

USCIS has explored AI-assisted translation capabilities in limited contexts. The issue raised by such tools is less about accuracy in isolation than about traceability.

Public disclosures do not clarify whether AI-generated translations are preserved in administrative records, whether they are distinguishable from human translations, or whether applicants can later determine the origin of linguistic discrepancies. In proceedings where credibility determinations may hinge on precise wording, the inability to trace translation sources could complicate appellate review.

Automation Bias and High-Volume Decision Environments

Research in psychology and public administration has identified automation bias as a recurring phenomenon: individuals may give disproportionate weight to automated outputs, particularly when operating under time pressure or high workloads.

USCIS adjudicators often review complex applications containing extensive documentation under significant time constraints. In such environments, automated classifications or flags may function as cognitive anchors even when formally advisory only.

This risk is structural rather than personal. Systems designed to increase efficiency may inadvertently encourage adjudicator reliance on automated outputs unless governance mechanisms explicitly measure and counterbalance that tendency. DHS has not publicly released data regarding override rates or officer reliance metrics that would allow evaluation of this issue.

Transparency, Administrative Law, and the Reviewability Problem

Artificial intelligence introduces new challenges to traditional administrative law principles. Judicial review of agency action typically relies on the administrative record to determine whether decisions were arbitrary or capricious. When algorithmic tools influence how evidence is organized or prioritized, their role may not appear explicitly in that record.

This creates a transparency paradox. Full disclosure of algorithmic criteria could allow applicants to tailor submissions strategically, potentially undermining fraud detection efforts. At the same time, minimal disclosure limits the ability of applicants and reviewing courts to understand how decisions were reached.

Balancing enforcement integrity with procedural transparency may become one of the defining governance challenges of AI-assisted administration.

The Path Forward: Questions That Remain Unanswered

Artificial intelligence will likely remain part of immigration application processing. The relevant policy question is not whether AI should be used, but how its use should be governed and evaluated.

Several questions remain unresolved:

• What information do adjudicators see after AI processing occurs?
• How frequently are automated outputs overridden?
• When an issue is flagged by AI, what are adjudicators required to review?
• Are AI-generated analyses preserved or logged for later review?
• How are bias and error rates measured longitudinally?
• What independent auditing mechanisms exist for rights-impacting systems?

Answering these questions would not undermine efficiency. Instead, it would strengthen confidence in a system whose decisions profoundly affect individuals and families.

Conclusion: Modernization Without Visibility

USCIS’s adoption of artificial intelligence reflects a broader transformation across government. Agencies facing unprecedented scale increasingly rely on automation to manage workload. DHS deserves credit for publicly acknowledging AI deployment through its inventory. Disclosure of existence is not the same as disclosure of impact.

At present, artificial intelligence operates largely behind the scenes, influencing workflows in ways that applicants, attorneys, and courts may not fully see or understand. That reality does not establish unfairness. It does, however, make transparency, governance, and ongoing evaluation essential.

Immigration adjudication has always balanced efficiency against individualized review. As algorithmic tools become more embedded in administrative processes, ensuring that balance endures may become one of the central challenges of modern immigration law.

For assistance with your business or immigration status, please contact BMD Member Robert Ratliff at raratliff@bmdllc.com. With over 25 years of trial experience in criminal defense and immigration law, Robert’s unique insights as a former Immigration Judge allow him to offer strategic guidance for clients facing complex immigration challenges.