Who Should Have the Final Say? Human-in-the-Loop Border Control

At the heart of the debate over AI in border control lies a question more important than model accuracy: Who should have the final say? Even if a system operates with high technical accuracy, decisions regarding entry into the country, asylum, visas, and security have serious legal consequences for individuals. Human oversight is not merely a matter of placing an approval button on a screen; the officer must be able to understand, question, and override the decision.

Four Levels of Automation

Human-onlySimple toolsFull control and decision-makingHuman-in-the-loopEvidence and recommendationsFinal decisionHuman-on-the-loopStandard decisions are automaticSupervision and exceptionFully autonomousAll analysis and decision-makingPost-processing

A single level of border management is not suitable for all tasks. MRZ check digits or SOD signatures can be processed automatically. However, conflicting travel justifications or requests for international protection require human assessment.

Task-based automation matrix

| Görev                              | Önerilen seviye       |
|------------------------------------|-----------------------|
| MRZ check digit                    | Otomatik              |
| Passive Authentication             | Otomatik              |
| DG hash karşılaştırması             | Otomatik              |
| Yüz eşleştirme                     | Otomatik + eşik       |
| Morph/liveness şüphesi             | İnsan incelemesi      |
| Vize koşulu belirsizliği           | İnsan incelemesi      |
| Giriş reddi                        | Yetkili insan kararı  |
| İltica/korunma beyanı              | Uzman insan süreci    |
| Çocuk ve insani istisna            | İnsan değerlendirmesi |

Meaningful human oversight

The officer should not merely be a passive operator who simply approves the AI result. The system must ensure the following conditions:

• Sub-checks and supporting evidence must be visible.
• Model confidence and known limitations must be disclosed.
• The operator must be able to request additional documentation or a re-capture.
• The AI recommendation must be modifiable, and the reason for any override must be recorded.
• Sufficient time and training must be provided for decision-making.
• Passengers must have access to a channel for filing objections or requesting a reevaluation.

Risk of automation bias

Human oversight alone is not a guarantee of security. If an agent becomes accustomed to the system being correct most of the time, they may approve recommendations without questioning them. Conversely, too many false alarms can cause real risks to be overlooked.

def oversight_metrics(events):
    return {
        "override_rate": ratio(events, "HUMAN_OVERRIDE"),
        "review_to_clear": ratio(events, "REVIEW_CLEARED"),
        "review_to_deny": ratio(events, "REVIEW_DENIED"),
        "median_review_seconds": median_duration(events),
        "operator_disagreement": disagreement_by_operator(events),
        "appeal_reversal_rate": ratio(events, "APPEAL_REVERSED")
    }

A very low override rate may not indicate that the system is flawless, but rather that operators are overly reliant on the AI. Metrics should be interpreted in context and through case studies.

What should the decision screen look like?

{
  "recommendation": "MANUAL_REVIEW",
  "summary": "Belge geçerli; vize koşulu çözümlenemedi",
  "checks": [
    {"name": "Document signature", "result": "PASS"},
    {"name": "Face match", "result": "PASS", "scoreBand": "HIGH"},
    {"name": "Visa condition", "result": "UNRESOLVED"}
  ],
  "suggestedActions": [
    "OPEN_AUTHORIZED_VISA_SOURCE",
    "ASK_FOR_SUPPORTING_DOCUMENT"
  ],
  "prohibitedAction": "AUTOMATIC_DENIAL"
}

The screen should not merely display red or green; it should highlight the unresolved issue and the legal actions the operator can take.

Special and Sensitive Situations

Alternative processes must be in place for children, the elderly, people with disabilities, individuals with facial alterations, and passengers who cannot be biometrically matched for technical reasons. A passenger claiming asylum or seeking protection should be directed directly to an authorized specialist rather than having a risk score generated.

Appeal and Re-evaluation

Passengers should be able to understand which automated system contributed to the process and, in the event of a negative outcome, request a human reevaluation. Appeal records may be fed back into model performance metrics; however, legal case data should not be converted into training data without proper oversight.

Management and Accountability

For each agent and model, the system owner, data owner, operations owner, and final decision-maker must be identified. The use of third-party models does not absolve the public authority of its responsibility. Decisions regarding deployment, version upgrades, and decommissioning must be documented.

Conclusion

The most appropriate goal in border control is not to remove humans from the loop, but to automate standardized and repeatable controls, thereby allowing humans to focus on legal judgment and exceptions. Meaningful human oversight involves the combined presence of authority, timing, justification, override, fallback, and appeal mechanisms.

References

• NIST AI RMF — Human-AI Interaction
• EU AI Act
• Frontex — Fundamental Rights