No expectant moms at the border: Trump's birthright Plan B - Axios

The Supreme Court's recent decision overturning birthright citizenship restrictions sent shockwaves through Washington. But for engineers and technologists, the real story lies in the unspoken infrastructure that would have been required to enforce it. Trump's "Plan B" - as reported by No expectant moms at the border: Trump's birthright Plan B - Axios - would have demanded a data architecture far beyond anything currently deployed at Customs and Border Protection. The policy's failure wasn't just legal; it was deeply technical.

What happens when immigration policy meets machine learning-and fails, This isn't merely a political storyIt's a case study in how poorly designed software systems can amplify bias, violate privacy. And create unenforceable mandates. For those of us who build systems at scale, the "birthright Plan B" debacle offers hard-won lessons about data integrity, algorithmic fairness, and the limits of real-time decision-making.

The Data Problem: Identifying 'Expectant Moms' at Scale

At the core of the proposed policy was a deceptively simple requirement: determine whether any woman crossing the U. S border was pregnant. And then classify her status for birthright citizenship eligibility. In a single port of entry, CBP processes thousands of travelers daily. Implementing such a policy would require real-time medical inference without traditional diagnostic tools - a task that current facial recognition and thermal imaging systems can't accomplish reliably.

CBP's current biometric screening uses facial recognition to match travelers against watchlists and visa databases. These systems, based on convolutional neural networks like those in TensorFlow and OpenFace, can identify individuals with ~99% accuracy in controlled conditions. However, detecting pregnancy from a two-dimensional image or thermal signature is fundamentally unsupported by any peer-reviewed computer vision model. A 2021 study by MIT Media Lab found that no commercial deployment of facial recognition can reliably infer pregnancy, let alone gestational stage. The policy's technical foundation was therefore built on a fantasy.

Machine Learning Models for Predicting Border Crossings: A House of Cards

Even if direct detection were impossible, the proposed "Plan B" might have relied on predictive analytics to flag women likely to give birth soon after crossing. The Department of Homeland Security already employs risk-scoring algorithms - such as the Automated Targeting System - to assign threat levels to travelers. These models ingest hundreds of variables: travel history, financial data, social media activity. And even behavioral cues from interrogation.

But extending these models to predict pregnancy status crosses a dangerous ethical line. Training a classifier to predict pregnancy would require a labeled dataset comprising ground-truth medical information - which DHS doesn't possess legally or ethically. Without such data, any model would become a proxy for other correlated variables: age, gender, country of origin. And even clothing style. This approach guarantees systemic bias. An analysis of the DHS PRIDE system (used for immigration benefit adjudication) revealed that machine learning models often magnify existing disparities in enforcement, especially against women from Central America.

Furthermore, the black-box nature of these models violates both the Algorithmic Accountability Act and basic software engineering principles. Without interpretability, no engineer can verify that the model isn't making decisions based on protected attributes like race or national origin - both of which are illegal under the Immigration and Nationality Act.

The Infrastructure That Failed: A Software Engineering Post-Mortem

From a pure engineering perspective, the birthright Plan B would have demanded a data pipeline capable of ingesting, processing. And disseminating pregnancy-related alerts across dozens of ports of entry in near real-time. That infrastructure doesn't currently exist at CBP. A 2023 Government Accountability Office report found that over 60% of DHS major IT acquisitions have suffered cost overruns or schedule delays. The agency's core case management system, TECS (Treasury Enforcement Communications System), was originally built in the 1970s and still runs on COBOL in parts.

Imagine the engineering effort: you need to create a secure, HIPAA-compliant database to store pregnancy status (which is protected health information), integrate it with CBP's existing traveler databases. And then expose real-time APIs to field agents' mobile devices. You also need robust access controls, audit logs, and failover mechanisms. Any government contractor who has worked on DHS projects will tell you such an undertaking would take at least five years and cost upwards of $2 billion - far beyond the timeline of an executive order.

Moreover, the system would need to handle edge cases: false positives (non-pregnant women flagged incorrectly), privacy litigation from ACLU, and the inevitable data breaches. As The New York Times reported, some legal scholars warned the policy was "surprisingly close" to constitutional violations - but the technical infeasibility was rarely discussed.

Facial Recognition at the Border: How It Works (and Why It's Controversial)

CBP's facial recognition program is already operational at over 200 airports and seaports. When a traveler approaches the gate, a camera captures their face, converts it to a mathematical template. And compares it against a gallery of visa photos and passport images. The system uses a 1:1 verification (matching you to your identity) or a 1:N identification (searching against a watchlist).

The technology relies on deep neural networks like FaceNet or ArcFace. These models achieve high accuracy under controlled lighting and pose, but at the border - with variable sun exposure, hats, masks, and expressions - error rates climb. A 2020 study by the National Institute of Standards and Technology found that facial recognition algorithms have higher false-match rates for Asian and African-American individuals compared to white individuals. For a policy targeting predominantly Hispanic pregnant women, the bias would be catastrophic.

Furthermore, the scale of data collection is staggering. CBP claims it stores face vectors temporarily (12-48 hours) for most travelers. But due to a patchwork of legacy systems, retention may extend far longer. A whistleblower in 2022 revealed that CBP's Enterprise Data Warehouse sometimes retains biometrics for years without proper oversight. The combination of unreliable detection and unchecked surveillance is a recipe for civil rights violations.

The Role of Real-Time Data Integration: APIs, Data Lakes. And Privacy

Enforcing birthright restrictions would require connecting CBP's travel databases with health records from the Centers for Disease Control (CDC) and state-level birth registries. This type of data integration is a nightmare of distributed systems engineering. Government agencies rarely expose production-grade APIs; instead, they rely on batch file transfers via SFTP, resulting in data that's hours or days old. Even within DHS, the data lakes used by ICE, CBP. And USCIS are notoriously siloed, each with its own schema and security model.

A real-time alert system would demand an event-driven architecture using something like Apache Kafka or Amazon Kinesis to stream traveler arrival events, then trigger a machine learning inference pipeline. Then the system must respond within seconds to a field agent's mobile query. And all of this must comply with the Privacy Act of 1974. Which prohibits collecting information on how individuals exercise their First Amendment rights (including pregnancy status as a protected health decision).

The integration challenge becomes even more daunting when you consider that most ports of entry lack reliable high-bandwidth internet. In remote border crossings along the Rio Grande, agents rely on cellular hotspots with limited coverage. The system must therefore support offline-mode capabilities with eventual consistency - a requirement that would dramatically increase complexity and testing requirements.

Ethics of AI in Immigration: The Algorithmic Bias Problem

Any system that attempts to infer pregnancy from data will inevitably rely on proxies: shopping patterns (purchases of maternity clothes or prenatal vitamins), travel to countries with medical tourism, or even changes in gait biometrics from surveillance cameras. Each of these proxies introduces bias. A low-income woman may not have digital purchase records; an undocumented woman may avoid formal medical care. The model would thus systematically miss certain populations while falsely flagging others.

Studies from the AI Now Institute and the Algorithmic Justice League have repeatedly shown that predictive policing and immigration enforcement algorithms disproportionately impact communities of color. In 2021, Amazon's Rekognition was shown to falsely match 28 members of Congress to mugshot databases, many of whom were people of color. If a large tech company can't even reliably match faces, expecting a government system to correctly infer pregnancy status is naive - and dangerous.

Beyond bias, there is the issue of consent. Travelers crossing the border have reduced Fourth Amendment protections. But collecting intimate biometric data for health inference without explicit consent likely crosses ethical lines. The IEEE Global Initiative on Ethics for Autonomous and Intelligent Systems explicitly advises against using AI to infer sensitive attributes without informed consent. This policy would have violated every major ethical AI framework.

Could Blockchain Solve Birthright Verification? A Thought Experiment

Some technologists have proposed using a blockchain-based self-sovereign identity (SSI) system to manage citizenship verification. Under this model, each individual would hold a digital wallet containing verifiable credentials (e g., a birth certificate issued by a hospital). Border agents could then cryptographically verify the credential without needing a centralized database of pregnancy status. This approach would theoretically enhance privacy and reduce the risk of surveillance.

However, the practical hurdles are immense, and the US healthcare system lacks a standardized digital identity framework. Most births are still recorded using paper certificates that are later digitized into state-level Vital Records Systems. Which use varying data formats (HL7 v2 vs. FHIR R4, for example). Connecting these systems to a blockchain requires universal adoption of a credential format like W3C Verifiable Credentials. Which is still not widely deployed. Moreover, blockchain solutions are notoriously slow for real-time verification and create new attack surfaces for identity theft. As a 2023 NIST report concluded, blockchain for identity is "promising but not yet production-ready for critical government functions. "

What Developers Can Learn from This Policy Battle

The birthright Plan B saga reveals three critical lessons for engineers working on government or regulated systems. First, know the domain. Building a system to enforce a policy you don't fully understand - especially one with medical and legal implications - is reckless. Every engineer should demand clear, unambiguous requirements from domain experts. And push back when those requirements are technically impossible.

Second, test for bias from day one. If your training dataset lacks representation, your model will fail marginalized groups. Use fairness metrics (e, and g, demographic parity, equal opportunity) as KPIs, not afterthoughts. Integrate TensorFlow's Responsible AI Toolkit or the IBM AI Fairness 360 library into your CI/CD pipeline.

Third, architect for reversibility. Good software allows policy changes to be rolled back gracefully. Use feature flags to disable new enforcement logic without redeploying; maintain audit trails that can support after-the-fact accountability. The best way to prevent abuse is to build a system that every action is logged, reviewable. And revocable.

Frequently Asked Questions