EU AI Act: High-Risk AI in Migration, Asylum and Border Control (Annex III Point 7)

Overview: AI at the Border Under Strict Rules

Migration, asylum and border control management is the seventh high-risk area in Annex III of the EU AI Act (Regulation (EU) 2024/1689). The recitals explain why: people in migration procedures are in a particularly vulnerable position, dependent on the outcome of decisions by authorities, and errors by AI systems in this context can affect non-refoulement, the right to asylum, private life and the rights of the child. At the same time, European borders are among the most digitised in the world, with biometric databases, risk-screening systems and automated document checks already embedded in daily operations. Point 7 brings that machinery under binding quality, transparency and oversight requirements.

What Annex III Point 7 Covers

The provision applies to AI systems intended to be used by or on behalf of competent public authorities, or by Union institutions, bodies, offices or agencies, in the field of migration, asylum and border control management, insofar as use is permitted under applicable law. Four use cases are listed:

• 7(a) Polygraph-type tools: AI systems used as polygraphs or similar tools
• 7(b) Risk assessment: AI systems used to assess a risk — including a security risk, a risk of irregular migration or a health risk — posed by a natural person who intends to enter or has entered the territory of a Member State
• 7(c) Application processing: AI systems used to assist competent authorities in the examination of applications for asylum, visa or residence permits and of associated complaints with regard to eligibility, including related assessments of the reliability of evidence
• 7(d) Identification: AI systems used for detecting, recognising or identifying natural persons in the context of migration, asylum or border control management, with the exception of verification of travel documents

The travel document exception in 7(d) is practical: a kiosk that checks whether a passport chip matches the face of the person presenting it performs one-to-one verification and is excluded from this point, whereas systems that search for or identify persons against databases are included.

Interaction With Bans and Other Annex III Points

Border AI must first be screened against Article 5. Emotion recognition is prohibited only in workplaces and education, so it is not banned at borders — but an AI lie detector that infers emotional states would be high-risk under both 7(a) and the biometrics rules, and faces transparency duties under Article 50(3). Real-time remote biometric identification prohibitions in Article 5(1)(h) apply to law enforcement purposes; identification at borders for migration management purposes is instead governed by point 7(d) as high-risk. Where border authorities act in a law enforcement capacity, Annex III point 6 may apply in parallel. Classification memos in this field should map each system against points 1, 6 and 7 together.

Who Must Comply

Providers include defence and security contractors, biometric technology vendors, and increasingly the EU agencies and national ministries that develop systems in-house — a public body that puts an AI system into service under its own name is a provider. Deployers include national border guards, immigration and asylum agencies, consulates processing visas, and Union bodies such as the agencies operating in this field. Several reinforced duties apply because deployers are public authorities: a fundamental rights impact assessment under Article 27 before first use, registration of deployment in the EU database under Article 49 (in the secure non-public section for this domain, per Article 49(4)), trained human oversight under Article 26, and information duties towards affected persons where high-risk AI supports decisions about them. Article 74(8) routes market surveillance to data protection authorities or equally independent bodies.

The Large-Scale EU IT Systems Transition

One transition rule is unique to this field. Article 111(1) provides that AI systems which are components of the large-scale IT systems listed in Annex X — the Schengen Information System, Visa Information System, Eurodac, the Entry/Exit System, ETIAS and the interoperability framework — and which were placed on the market or put into service before August 2, 2027, must be brought into compliance by December 31, 2030. New systems and significant modifications fall under the standard regime. National systems outside Annex X follow the normal August 2, 2026 deadline for Annex III classifications, so authorities should not assume the 2030 horizon covers their whole estate.

Practical Compliance Steps

1. Inventory every AI component across visa processing, asylum case handling, border checks, risk screening and surveillance, including modules embedded in vendor platforms
2. Map each system against Annex III points 1, 6 and 7 and against Article 5, documenting the conclusions
3. Identify which systems are Annex X components with the 2030 transition and which face the 2026 deadline
4. Complete fundamental rights impact assessments with specific attention to non-refoulement, child rights and discrimination risks
5. Validate accuracy across demographic groups: error rates in face recognition and risk scoring must be tested on populations actually encountered at borders
6. Train caseworkers and border officers to understand system limits, override outputs and record the reasons

Concrete Examples

Example one: a consulate uses software that triages visa applications by predicted risk of overstay, flagging cases for deeper review. That is point 7(b) and arguably 7(c) — high-risk; the human caseworker remains responsible, and the system needs documented bias testing because nationality-correlated features can produce discriminatory outcomes.

Example two: a border agency deploys cameras with face recognition to identify persons crossing outside official border points. That is identification under 7(d) — high-risk, with data protection law applying cumulatively.

Example three: automated passport gates that verify a traveller against the chip in their own document perform verification of travel documents and are excluded from 7(d), though other law still applies.

Action Before August 2, 2026

Authorities and their vendors should treat the remaining months as delivery time, not analysis time. Procurement frameworks need AI Act conformity clauses now, because border technology contracts run for years. Vendors should complete Annex IV technical documentation and conformity assessment for systems already in national deployments. Agencies should sequence fundamental rights impact assessments before the deadline rather than after, and align AI logging with existing Schengen evaluation and data protection audit mechanisms. With penalties reaching 15 million euros or 3 percent of turnover for high-risk violations, and the political sensitivity of border technology, this is one of the areas where supervision is expected to be active from day one.

Fundamental Rights Anchors: What Assessments Must Address

The fundamental rights impact assessment under Article 27 is not a formality in this domain, because the rights at stake are among the strongest in the EU legal order. Assessments should explicitly address non-refoulement — the prohibition on returning a person to persecution or serious harm — since an AI error in asylum evidence assessment can contribute to exactly that outcome. They should address the rights of the child, given that minors appear in migration flows and that age assessment technologies are themselves contested. They should address non-discrimination, because risk screening features correlated with nationality, language or travel patterns can encode prohibited distinctions. And they should address effective remedy: Article 86 of the AI Act gives persons affected by decisions based on high-risk AI output the right to a clear and meaningful explanation, which in this field must be deliverable to people who may not speak the procedural language and may be represented by counsel only later, if at all.

Operationally, this means assessment documents should name the specific decision points where AI output enters a case file, the safeguards at each point, the escalation route when a caseworker doubts the output, and the mechanism for correcting database entries when an identification or risk score proves wrong. Cross-border data quality is a recurring failure mode: an erroneous flag created in one Member State propagates through shared systems, so correction procedures must work across borders too. Authorities that build these answers into their assessments before August 2026 will not only comply but materially reduce the litigation exposure that has historically followed border technology deployments.

For vendors, the commercial implication is clear: the European border technology market now has a single, binding quality bar, and conformity documentation is becoming the entry ticket to public tenders. Suppliers from outside the EU should appoint their authorised representative early, prepare technical documentation in the Annex IV structure rather than internal formats, and expect customers to request demographic performance evidence as standard. Suppliers that already serve regulated identity markets will recognise most of the discipline; the genuinely new elements are the fundamental rights framing, the EU database registration and the explanation rights of affected persons, all of which need to be designed into products rather than promised in proposals.

This article is for informational purposes only and does not constitute legal advice. Regulatory requirements change frequently — verify current rules with official sources. Built by Sawai Gyoseishoshi Office, Hiroshima, Japan.