At 6:40 in the morning, the border line is not really a line. It is a queue of faces waiting to become numbers.
You walk up, a camera looks at you for about a second, and software decides whether the face in front of it matches the face on a document. A human waves you through. Nobody says what the machine thought.
Most people miss the part that is not about algorithms at all. The experience is about silence. The short answer: the border is the most automated part of the AI state, and it has the weakest appeal of any system that decides your future. Three machines will touch a normal trip, and one is measurably wrong about children by years.

Checkpoint One: The Camera You Do Not Get to Argue With
On October 5, 2026, US Customs and Border Protection announced that its facial recognition system had verified the identity of more than one billion travelers.
This is not a pilot. CBP says the system runs at entry in every US airport with international arrivals, at exit in 66 airports and 42 seaports, and in all pedestrian lanes at land crossings. Congress added $673 million to expand it to every port of entry. The agency credits it with catching more than 600,000 overstayers and more than 2,300 impostors since 2018.
The design is the story:
- US citizens are not covered by the rule, but they get scanned anyway unless they speak up. CBP says citizens can decline and have their passport inspected manually.
- Citizen photos are deleted within 12 hours. Good policy, not a law you can enforce at the kiosk.
- The opt-out is quiet. A 2022 GAO report found CBP’s privacy signage gave “limited information on how travelers could request to opt out” and was “not always posted.”
A face match is a claim about identity. The person carrying the face cannot inspect the claim or point to an error rate.
Checkpoint Two: Your Visa File Gets Read Before a Person Sees It
The second machine does not look at your face. It reads your paperwork.
USCIS now runs machine learning across filings before an adjudicator opens them. The clearest example is the ELIS Evidence Classifier, which tags uploaded evidence and decides which documents an officer sees first. According to a Cozen O’Connor alert published April 27, 2026, the system has processed more than 24 million page scrolls, and USCIS has published no error-rate data.
Lawyers are reporting the symptoms: Requests for Evidence demanding documents that were in fact submitted, the signature of a mis-tagged exhibit. A Verification Match Model cross-checks names and dates of birth against E-Verify and SAVE, where a small inconsistency can trigger a flag. The systems also read hidden text layers, so leftover template text in a PDF can produce a deficiency claim that looks nonsensical to the applicant and rational to the classifier.
The aggregate numbers are moving. Cozen reports EB-2 NIW denial rates near 40 percent, with AI-assisted screening cited as a driver of rising RFEs and Notices of Intent to Deny. It also cites the case that should be a warning label: more than 1,200 international students lost legal status because of automated database terminations, which took nationwide litigation and a government reversal to fix.
A database mismatch ended legal status for over a thousand people. The remedy existed. It required a federal lawsuit.
CBP also runs Babel X, an AI tool that scans social media and open-source data at ports of entry. Visa applicants must list five years of social media identifiers. A joke from 2021 becomes evidence in 2026.

Checkpoint Three: The Age Scan, Where the Errors Are Measured in Years
The third machine has the worst numbers, and it is aimed at children.
In November 2025, the UK Home Office announced it would pilot facial age estimation on asylum seekers, framing it as a way to “root out illegal migrants gaming the system by posing as children.” By May 2026 it had contracted Cognitec Systems to supply the algorithm. Lighthouse Reports, working with WIRED and The Independent, obtained the Home Office’s own internal test report, dated April 2025. The report tested seven algorithms against 2.5 million images.
What the Home Office’s own testing found:
- The best performing vendor overpredicts the age of minors. On average, a 17-year-old is predicted to be over 18.
- It performs worse on females, and much worse on Sub-Saharan Africans. For male Sub-Saharan Africans the average error rate was double that of other groups. For female Sub-Saharan Africans under 18, the average error was 4.6 years. A 14-year-old could be read as an adult.
- The bias lands on the biggest group. In 2025, the largest group of asylum seekers with age assessments raised were from Sub-Saharan Africa, per Home Office data. The system performs worst on the people it is most likely to be used against.
Lighthouse then analyzed NIST’s independent benchmarks for Cognitec, submitted April 2026, and calculated what the public report cards do not print. On NIST’s lower-quality “border” dataset, which resembles a first-encounter photo taken under time pressure, the system misclassified more than two thirds of 16-year-olds as adults. Even on the clean “application” dataset, more than a third were misclassified. More than half of 16-year-old West Africans were read as over 18, against less than a quarter of 16-year-old Eastern Europeans.
There is a fix, of a sort. Raising the “challenge age,” the threshold above 18 at which someone is treated as a minor, reduces the number of children read as adults. But at a challenge age of 24, a West African 16-year-old becomes more than 10 times more likely to be misclassified than an Eastern European 16-year-old. The tradeoff concentrates the bias instead of removing it.
The Home Office says the tool is “assistive,” “does not replace or overrule human judgement,” and that in cases of uncertainty people “will always be treated as children until a further assessment is conducted.” Cognitec said demographic differences “apply to all algorithms” on the NIST benchmark.
Why this matters most: a false airport match means secondary inspection. An age estimate decides whether someone gets child protections, child housing, a guardian, and a different legal track. The error is not a delay. It is a childhood.
The Lie Detector Nobody Can Validate
This one is not in production yet. That is its only reassuring feature.
Trouw reported in June 2026 that the Netherlands’ TNO and Amsterdam company Vicarvision ran a European Commission funded project from 2018 to 2021 testing whether AI could detect deception at border checks. It involved 25 organizations and drew nearly 8 million euros. Its effectiveness results are classified.
The science is not classified, and it is not kind. Bruno Verschuere, a professor of forensic psychology at the University of Amsterdam, told Trouw: “We know from hundreds of studies that the link is very weak. This time it is with AI, but in essence they keep doing the same thing in a different jacket.”
TNO and Vicarvision reject the label, saying the tool reads physical behaviors and only supports officers. But the EU’s AI Act treats AI-based lie detection as high risk and generally prohibits it, with narrow carve-outs for migration controls. When a technology is banned for citizens and permitted for migrants, the ban is about who is allowed to refuse.

The Pattern: A Score, a Signature, and No Appeal
The design repeats.
- The machine produces a flag. A non-match, a mis-tagged exhibit, an estimated age.
- A human signs off. True, and the weakest defense, because the human sees the flag and not the error rate behind it.
- The person is not told. You learn something went wrong when a visa is refused, not when the classifier misfired.
- There is no error rate to argue with. USCIS has not published one. The Home Office report was leaked, not released. The deception results are classified.
- The remedy is litigation or luck. The 1,200 students got their status back because lawyers forced the issue.
Borders have always been a place where ordinary protections thin out. What is new is the scale and the silence: a system that can judge a billion people and explain itself to none of them. The EU’s Entry/Exit System, fully operational since April 10, 2026, replaces passport stamps with facial images and fingerprints for non-EU short-stay travelers.

What Actually Works
There is no clean opt-out. There are specific moves that change outcomes.
US citizen at a US airport: you can decline the face scan. Tell the officer or airline representative you want a manual passport inspection. CBP’s own statement says this is allowed.
Flying to Europe: EES biometric capture is mandatory for non-EU short-stay travelers. Build extra time into your first entry.
Filing pending: assume a machine read it first, and keep an exact copy of what you submitted. If you get an RFE for a document you know you sent, do not resend blindly. Cite the exhibit, the page, and the timestamp. Check PDFs for hidden text layers. Those artifacts are now evidence.
In an age dispute: ask two questions in writing. What threshold is being applied, and did the tool inform the decision or make it? The Home Office’s position is that uncertainty should resolve toward treating someone as a child. Hold them to it, and get a lawyer early.
Want the numbers: file a public records request. That is how the Home Office report surfaced.
The Cameras Do Not Stop at the Border
Face recognition and license plate readers are already in your grocery store parking lot, your downtown, and your neighbor’s doorbell. The border is just the version with a legal exemption built in.
FAQ
Does the US scan every traveler’s face at the airport?
At airports with international arrivals, facial comparison runs at entry, and at exit in 66 airports. US citizens can ask for a manual inspection. Noncitizens are covered by the biometric rule finalized in 2025.
Can I refuse a face scan?
US citizens can, by telling an officer or airline representative and asking for manual document inspection. Noncitizens at US ports of entry, and travelers entering the EU under EES, generally cannot.
Is a human making the final decision?
Agencies say yes, and formally that is true. In practice the human reviews a flag from a system whose error rate has not been published.
What if an automated flag hits my case?
Get the record in writing. Ask what system flagged the file, what data it used, and whether a person reviewed it. Then answer the specific deficiency, not the whole filing.
The same scoring logic follows people past the checkpoint and into the workplace, which is what our guide to workplace AI surveillance covers.
The Short Version
- The border is the most automated part of the AI state. CBP has verified more than one billion travelers by face, and the EU now fingerprints and photographs non-EU short-stay visitors.
- US citizens can opt out and ask for manual passport inspection. Most do not know that.
- Your visa file is screened first. USCIS’s evidence classifier has processed over 24 million page scrolls and published no error rate.
- The UK’s asylum age scanner is the clearest failure. Its own tests read a 17-year-old as an adult on average, with a 4.6-year average error for Sub-Saharan African girls under 18.
- On NIST border-quality photos, more than two thirds of 16-year-olds were classified as adults. The government proceeded anyway.

