Human Review of AI-Assisted Asylum Decisions A Procedural Reading of the AI Act and EU Asylum Law

Abstract

Asylum authorities may keep the formal act of decision-making in human hands while allowing digital systems to shape the case long before signature. Speech recognition, translation, retrieval, credibility flags and triage tools can decide which facts are visible, which doubts are treated as important and which procedural route a claim follows. This article argues that the distinction between human and automated decision-making is therefore insufficient. The central issue is whether a person affected by an AI-assisted asylum process can understand and challenge the role played by the system before the decision becomes legally fixed. Reading the EU AI Act together with the Asylum Procedure Regulation, the Qualification Regulation, the GDPR and the Charter, the article proposes a standard of reviewable decision provenance. That standard requires a traceable account of how a relevant output was produced, used, contested and translated into reasons. It does not demand disclosure of every technical operation. It requires enough information for the authority, the applicant and a court to test whether the legal assessment remained independent, participatory and reviewable.

Keywords

artificial intelligence; international protection; decision provenance; human review; effective remedy

1 Introduction

In an asylum file, the decisive digital intervention may occur before the refusal or grant is drafted. An automated transcript can turn hesitation into certainty; a retrieval tool can place one country report ahead of competing material; and a language or document system can make one doubt appear more relevant than others. None of those operations formally decides the claim. Yet each may organise the factual world within which the responsible officer later works.

The EU Artificial Intelligence Act adds an important regulatory layer for migration and asylum technologies. Certain systems used to assist the examination of protection claims, assess evidence or support border-control functions are classified as high-risk. That classification brings duties concerning risk management, data governance, transparency, human oversight and logging. The Act, however, does not itself supply the full procedural answer for asylum law.

This article argues that oversight must be understood procedurally, not merely technically. Article 14 of the AI Act concerns whether a system is designed so that a person can recognise limitations and intervene. Article 26 concerns how the deployer uses the system in practice. EU asylum law asks a further question: whether the applicant still receives an individual examination, meaningful participation, a reasoned decision and an effective remedy.

The article develops the concept of reviewable decision provenance. Provenance, in this sense, means a legally sufficient account of an influential output: what material entered the system, what operation was performed, who received the result, how it affected the case, whether it was challenged and how the authority assessed it. The aim is not complete technical disclosure. The aim is to prevent an automated contribution from becoming an untested premise inside a human decision.

2 Machine influence before the legal conclusion

2.1 Producing the evidentiary record

An official file is never a neutral container. It is built through choices about capture, translation, extraction and compression. Automated transcription illustrates the problem. A recording may contain hesitation, clarification, overlapping speech and an interpreter’s reformulation. The presented text may smooth those features away. If the official reads only the transcript, the applicant’s account may already have been narrowed.

The procedural consequence is clear. Where automation changes the representation of evidence, the transformed item should remain linked to its source. Review must be directed not only at the final reasons but also at whether the evidentiary record used for those reasons was lawfully and reliably produced.

2.2 Directing attention and assigning weight

A second group of systems does not create evidence but guides attention. Consistency detectors, document-authenticity tools and language-analysis systems may generate warnings, rankings or probabilities. Their output can look precise even when it rests on uncertain assumptions. The danger is that the authority begins from suspicion and asks the applicant to disprove a warning before deciding whether that warning has legal value.

Digital tools can also influence a file through silence. A search system may fail to retrieve a relevant source, or a detector may fail to flag a vulnerability. If review focuses only on visible warnings, it misses the way absence can shape the material actually considered by the decision-maker.

2.3 Chained outputs and feedback

Asylum technologies are often examined as single products, although files may pass through several connected tools. A transcript may feed a contradiction detector; the flagged passages may influence a later interview; and a drafting assistant may then convert the exchange into reasons. Each stage can make the earlier automated step appear more authoritative than it was.

For that reason, legal analysis should focus on the decision chain. For every materially influential output, the authority should be able to identify the source material, the transformation performed, the recipient of the output, the procedural consequence and any later dependency. This is not a demand for limitless documentation. It is the minimum record needed to decide whether human review had legal substance.

3 The allocation of responsibility in EU law

3.1 High-risk classification and material influence

The AI Act treats several migration, asylum and border-management functions as high-risk. Annex III point 7(c) covers systems used to assist public authorities in examining applications for asylum and related complaints, including assessment of evidence reliability. The provision reflects the seriousness of the context: a technical step in a protection case may affect the risk of removal, detention or loss of status.

For high-risk systems, provider obligations supply part of the infrastructure for responsible use. Article 13 requires instructions that allow deployers to understand capacities, limits, accuracy and foreseeable risks. Article 14 requires design measures supporting human oversight. These duties do not decide asylum claims, but they create the informational conditions without which officials cannot exercise judgment responsibly.

Deployers carry separate duties. Article 26 concerns use according to instructions, suitable staff, monitoring, log preservation and notice where a person is affected by a decision made or assisted by an Annex III system. Article 27 also requires certain public bodies to assess fundamental-rights impact before deployment. These rules matter because they connect the system’s technical operation with the authority’s public-law responsibility.

3.2 The asylum instruments: responsibility for evidence and procedure

The Asylum Procedure Regulation organises responsibility around the determining authority, not around software suppliers or administrative convenience. Its guarantees include information, counselling, attention to special procedural needs and a personal interview within the legal examination of the claim. These stages are not simple data collection. They are the means through which the applicant can influence the file before it is closed.

Article 34 places the substantive assessment on the authority itself. The official must consider statements, documents, current and reliable country information and the individual circumstances of the applicant objectively and impartially. A risk indication, authenticity warning or automated summary may assist that work, but it cannot become the legal assessment.

The same point follows from reason-giving and appeal. Article 36 requires written reasons for rejection, and Article 67 requires an effective first-instance remedy involving a full review of both factual and legal issues. Those guarantees lose force if an automated contribution has shaped the result but remains invisible. The file must therefore show how the automated material entered the reasoning and what weight the official gave it.

3.3 Participation, data protection and judicial method

The Court’s judgment in SCHUFA shows why formal labels are unreliable. An automatically generated score may itself amount to a decision where another body gives it decisive practical effect. The case does not transform every asylum tool into an Article 22 GDPR decision. Its value lies elsewhere: an intermediate output cannot be isolated from legal responsibility simply because a human being later signs the decision.

The relevant legal sources work cumulatively. The AI Act helps make a system intelligible and governable; the APR and Qualification Regulation determine how a protection claim must be examined; the GDPR regulates the processing of personal data; and the Charter protects participation, equality and remedy. Compliance with one layer cannot cure failure under another.

3.4 Staggered application and present obligations

The transitional timetable for the AI Act does not pause asylum safeguards, data-protection duties or Charter rights. An authority cannot defend an uncorrectable transcript or an unexplained evidentiary inference by relying on a future technical-compliance date. In the transitional period, each safeguard must be identified accurately: some duties arise from the AI Act, while others already follow from asylum procedure, data protection and fundamental-rights law.

4 Reviewable decision provenance

4.1 Ability to interrupt the automated sequence

The first element asks whether the reviewer can identify the problem and alter the next procedural step. Legal competence concerns refugee status, subsidiary protection, evidentiary assessment and procedural guarantees. Technical competence concerns what a specific output legally signifies, the data on which it depends and its uncertainty. An official may understand the software in general but still fail to detect a defective result in a particular file.

A meaningful power to override therefore requires more than a visible button. The person responsible must have access to the relevant source material, enough information about the output’s limits and institutional permission to depart from it. If disagreement produces delay, extra approval or negative performance consequences, advice may function as an instruction in practice.

4.2 Contestability while correction still matters

The second element concerns timing. Information supplied only after refusal may support an appeal, but it cannot restore questions that were never asked or sources that were never considered. Notice must therefore arrive while the applicant can still affect the administrative assessment.

Useful notice should identify the contested proposition, not merely announce that technology was used. In a transcription dispute, the audio segment and transcript matter. In a document-warning case, the applicant needs to know which feature was treated as irregular and what weight the authority proposed to give it.

4.3 Independent appraisal under the governing legal test

The third element prevents a machine output from hardening into an unofficial presumption. The applicant should not have to prove that a system is generally unreliable before questioning a particular result. The authority must first decide whether the output is relevant, whether the correct material was processed and whether the method answers a legally meaningful question.

4.4 A record that connects output to reasons

The fourth element concerns institutional memory. High-risk systems should facilitate logging, and deployers must retain accessible logs for a minimum period. Separately, asylum procedure requires a record capable of supporting participation and appeal. These duties overlap when an automated output materially shapes the case.

Reasons should state the factual proposition taken from the tool and the official’s own assessment of that proposition. A general sentence that AI was used is too abstract; copying a score is not enough either. If the output made no material difference, the record should say so briefly. If it mattered, the reasoning should explain why.

4.5 Judicial access to the disputed contribution

The fifth element asks whether a court can reach the actual dispute. A remedy limited to the final wording may miss why a document was discounted, an interview was shortened or a country source was not considered. Access may vary with the allegation, but the reviewing body must be able to examine the material that shaped the administrative result.

5 Institutional design for credible review

Individual care cannot compensate for a defective system of work. Before deployment, an authority should map the decision chain: what enters the tool, what it produces, who receives the output, which later systems consume it and what legal or procedural consequence follows.

Procurement should secure the evidence needed for responsible operation. Contracts should provide access to technical documentation, version history, validation material, incident information and litigation records. They should also suspend or modify a system when serious doubts appear and require lawful disclosure to supervisory bodies or courts.

6 Conclusion

The legal significance of AI in asylum procedure cannot be assessed only by asking who signs the final page. Software may determine the representation of an interview, the evidence that becomes visible, the doubt treated as decisive and the procedural route followed by the claim. A human signature is not enough if those earlier choices remain inaccessible.

Reviewable decision provenance offers a more accurate test. An influential output must remain within a chain that can be interrupted by a competent official, contested by the applicant before decision, assessed independently under the governing law, connected to intelligible reasons and examined by a court. The necessary safeguards differ with practical influence, not with the contractual label attached to the system.

This approach allows useful technological assistance while preserving legal responsibility. Reliable transcription, careful retrieval and transparent organisation may improve asylum adjudication. The limit is reached when an output becomes an inaccessible layer of public authority. Human review has legal value only where the machine’s contribution remains identifiable and answerable throughout the procedure.

References and Bibliography

Article 29 Data Protection Working Party, Guidelines on Automated Individual Decision-Making and Profiling for the Purposes of Regulation 2016/679, WP251rev.01 (6 February 2018).

Beduschi A, ‘International Migration Management in the Age of Artificial Intelligence’ (2021) 9 Migration Studies 576.

Case C-277/11, M.M. v Minister for Justice, Equality and Law Reform and Others, EU:C:2012:744.

Case C-634/21, SCHUFA Holding (Scoring), EU:C:2023:957.

Charter of Fundamental Rights of the European Union [2012] OJ C 326/391.

European Union Agency for Asylum, Digital Transformation of Asylum and Reception Systems, Situational Update No 23 (11 September 2025).

European Union Agency for Fundamental Rights, Assessing High-risk Artificial Intelligence: Fundamental Rights Risks (Publications Office of the European Union 2025).

Regulation (EU) 2016/679 (General Data Protection Regulation) [2016] OJ L 119/1.

Regulation (EU) 2024/1347 on standards for qualification for international protection [2024] OJ L 2024/1347.

Regulation (EU) 2024/1348 establishing a common procedure for international protection in the Union [2024] OJ L 2024/1348.

Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence [2024] OJ L 2024/1689.

Vavoula N, ‘Artificial Intelligence (AI) at Schengen Borders: Automated Processing, Algorithmic Profiling and Facial Recognition in the Era of Techno-Solutionism’ (2021) 23 European Journal of Migration and Law 457.

Wachter S, Mittelstadt B and Floridi L, ‘Why a Right to Explanation of Automated Decision-Making Does Not Exist in the General Data Protection Regulation’ (2017) 7 International Data Privacy Law 76.

Zalnieriute M, Bennett Moses L and Williams G, ‘The Rule of Law and Automation of Government Decision-Making’ (2019) 82 Modern Law Review 425.