An asylum applicant sits down to explain why returning home may place them in danger. The account may cross years, borders and languages. Some events may be difficult to name; some dates may be approximate; a detail that appears late may be the most important one. A smooth conversational interface could ask every standard question, translate the answers and flag inconsistencies before a caseworker enters the room. That efficiency is precisely why the boundary must be clear.
Technology can prepare, support and document an asylum interview. It should not conduct the consequential interview, infer credibility from fluency or emotion, or turn narrative inconsistency into a risk score. The interview is not data extraction with a friendly front end. It is a legally structured encounter in which a responsible person must listen, adapt, clarify and remain answerable for how the account is understood.
What the interview is for
The personal interview is not simply an opportunity to collect fields missing from a form. For applications lodged in the European Union from 12 June 2026, the directly applicable Asylum Procedure Regulation gives the applicant an opportunity to present the elements needed to substantiate the claim and to explain missing elements, inconsistencies or contradictions. It requires privacy and confidentiality, an interpreter able to support appropriate communication, trained staff, attention to personal and cultural circumstances, and the possibility of legal assistance. The interviewer must be competent to take account of age, gender, vulnerability and special procedural needs.
Those requirements describe a reciprocal process. The authority does not merely receive a story; it creates conditions in which the person can give it. The interviewer may need to slow down, reformulate a question, recognise that a term has no direct equivalent, separate a misunderstanding from a contradiction, or pause when the applicant cannot continue. A lawyer may identify an issue the person does not know is legally relevant. An interpreter may need to clarify a linguistic ambiguity without becoming a witness or decision-maker.
A chatbot changes that relationship even if its questions sound empathetic. It works from a model of expected inputs and next turns. When an answer does not fit, the system must classify it, select a follow-up or route it as an exception. That machinery is useful for low-consequence preparation. In the interview itself, it risks converting ambiguity into a hidden machine choice before anyone with legal responsibility sees it.
Where technology genuinely helps
There is much to improve around the interview. Scheduling systems can offer suitable times, send reminders through safe channels and record accessibility or childcare needs. Secure portals can let applicants and counsel submit documents, see what the authority has received and correct contact details. Case-management tools can assemble a chronology without presenting it as established fact. They can identify untranslated material, prepare an interpreter with approved terminology and bring relevant country information to the officer’s attention.
Technology can also make reliable procedural information available before the meeting. The person should know the interview’s purpose, who will attend, the role of the interpreter, whether it will be recorded, how to request an interviewer or interpreter of a preferred sex where the law permits, and how to obtain legal help. This is a good place for multilingual information tools, provided the authoritative source remains visible and a human route is available when the wording is unclear.
During and after the interview, audio recording, secure transcription and document linking can improve the record. The current EU regulation requires audio recording and a thorough factual report or transcript, with routes to comment on incorrect translations, misunderstandings and factual mistakes. Speech technology may help produce a draft transcript, but a probabilistic transcript should not silently replace the recording or the applicant’s correction rights. The value lies in finding and reviewing passages more easily, not in manufacturing an unquestionable record.
The same division of labour applies to language. Machine translation may help the authority identify a document or prepare material for a qualified interpreter. It must not become the sole channel through which testimony is heard. As Alkemata examined in Can Machine Translation Carry a Legal Right?, legal communication needs a correction path and a person who can preserve ambiguity rather than resolve it invisibly.
Why inconsistency is not a score
Asylum decisions often depend heavily on an applicant’s account because documentary proof may be incomplete or unavailable. That makes inconsistencies important, but not self-interpreting. A changed date can arise from deception; it can also arise from translation, different calendars, fatigue, shame, a misunderstood question, the passage of time or the way distress affects recall. Culture shapes how people sequence events, describe family relations and respond to authority. Dialect and code-switching can alter meaning before the words reach the official record.
The EUAA Practical Guide on Evidence and Risk Assessment treats evidence assessment as cooperation between applicant and determining authority. It says the personal interview requires more than making time and space available: the authority must provide guidance, use appropriate questioning and create the right conditions to gather relevant elements. Its discussion of interview distortions includes multilingual communication through an interpreter and factors affecting memory and disclosure.
An automated inconsistency detector can still assist preparation. It might place two dates or place names side by side and show their exact sources. The responsible officer can then decide whether the difference is material and ask an open clarification. The system should not label the person “unreliable”, calculate a credibility probability or rank cases by narrative conformity. A score hides which discrepancies mattered, how they were weighted and which alternative explanations were considered. It also invites anchoring: once the number appears, the human interview can become an exercise in confirming it.
Fluency and visible emotion are even weaker foundations for coercive inference. A person may speak smoothly because the account has been repeated to several authorities, or haltingly because they are using a second language. Eye contact, pauses, tone and facial movement vary with culture, disability, fear and the interview setting. Technology that converts these signals into confidence, stress or deception labels does not create knowledge of whether an event occurred. It creates a new claim about the person, one they may be unable to inspect or contest.
The strongest case for automation
The counterargument is not trivial. Asylum authorities face workload pressure, interviews vary in quality, and trained officers and interpreters are scarce. A structured conversational system could ask mandatory questions consistently, work at any hour, reduce omissions and let applicants disclose information without facing an intimidating official. Human interviewers can be rushed, sceptical or inconsistent; keeping a person “in the loop” does not guarantee fairness.
Those weaknesses justify better tools and stronger supervision, but they do not establish that the machine should occupy the interviewer’s chair. Standardisation is valuable where the question is genuinely standard: confirming identity data, arranging an appointment, explaining procedure or checking that a file opens. The consequential account is different because relevance emerges through interaction. A trained interviewer must judge when to depart from the script, when not to press, when to seek specialist support and when an apparent contradiction requires the authority to examine its own translation or records.
The better use of automation is to improve the human interview’s conditions. It can show the officer which documents remain unread, protect time for preparation, suggest topics without assigning conclusions, and record whether required safeguards were offered. Quality assurance can sample interviews, check whether applicants were allowed to explain adverse points and identify patterns across officers. None of this requires a system to impersonate understanding.
A legal boundary, not a user-interface preference
The EU’s consolidated AI Act places systems used by public authorities to assist examination of asylum applications, including assessments of the reliability of evidence, in its high-risk list. It separately includes polygraph-like tools used in migration and asylum. The Act generally applies from 2 August 2026, while the detailed obligations for Annex III high-risk systems are scheduled to apply from 2 December 2027. Classification is not permission, and compliance will not make every use appropriate. A technically documented credibility system could still be the wrong institutional design.
The Asylum Procedure Regulation sets a more concrete human boundary. Interviews are conducted by staff of the determining authority; applicants respond in person; officers require training and competence; interpreters and legal advisers have defined roles. Remote interviews may be justified, but they remain interviews between accountable participants. The EUAA’s 2025 remote-interview guidance treats video, audio, privacy, legal assistance and interpretation as conditions that must be actively managed, not as functions that disappear into software.
The authority must remain present
A human-centred asylum system should therefore draw the line by consequence. Technology may organise access, reduce clerical delay, protect the record and help professionals see relevant material. It should stop before it asks for the account on which protection may depend, interprets the person’s behaviour, or converts uncertainty into a recommendation about credibility.
That boundary does not romanticise human judgment. It makes responsibility locatable. The applicant should know who conducted the interview, who interpreted it, which record will be used, how to correct it and who will decide. Counsel must be able to challenge both factual treatment and technological assistance. The officer must have authority to disregard an automated prompt, investigate an error and explain the reasoning without borrowing the system’s confidence.
If you want to follow Alkemata’s continuing examination of public technology and human agency, return for future articles on where digital assistance strengthens institutions—and where restraint is the more capable design.
The remaining question for an asylum authority is not whether conversational technology can ask plausible questions. It is whether the institution can guarantee that no model, interface or procurement claim stands between a person’s account and the trained human being who must hear it, test it fairly and answer for the consequences.
