For the professional class, AI mostly means convenience. We open the chat, ask something, write an email to the plumber, or vibe-code an app over the weekend. If the model hallucinates or the generated code misbehaves, we sigh, close the laptop, and perhaps go out for a glass of natural wine.
However, for many others, especially people with little institutional power, such as asylum seekers, AI can have a more bitter flavour. Many of these people don’t get to decide whether they want to use such systems. They are subjected to them.
What is the difference between these two positions? One, of course, is the magnitude of the consequences. Minor annoyance or disappointment when the model fails to fulfil a request is very different from facing a situation in which your legal status in the country is at risk because of an error in the process. But there is also another difference.
It is about agency: the privilege of choosing whether to use AI or to opt-out. Some people have it, and many others don’t.
The Bundesagentur für Arbeit is using BAKIRA, a generative-AI research assistant. The system can summarise documents, retrieve information from the agency’s knowledge base, translate texts into more than 40 languages and help with other internal tasks.
This seems like a version of AI we have been promised: a helper that removes tedious work while leaving judgement to a human. The agency describes BAKIRA as an internal assistant rather than a system for making any decisions, but it is one of the signs of rapid AI adoption across different government agencies.
Lower Saxony’s justice system is developing similar assistants, including one designed specifically for asylum cases. Since 2024, the state has been working on EMIL, a chatbot designed to help courts research the information about countries of origin used in asylum proceedings. According to the Lower Saxony Ministry of Justice, EMIL searches through a collection of about 1.5 million documents, finds relevant pieces, summarises material and translates foreign-language sources.
Again, EMIL does not decide whether an asylum seeker will be allowed to stay in Germany or not; it is an assistant to a judge.
But even if these systems are just assistants, we cannot shrug off the question of their reliability. The quality of an AI-assisted search depends on the completeness and freshness of its source material, the system’s ability to retrieve the right documents, and its capacity to create meaningful numerical representations of them. Searching and synthesising information from a collection of 1.5 million documents is not a trivial task from a technical standpoint.
Legal-tech companies are working, often with substantial investment, on very similar problems. But their relationship to the technology is different. A law firm can decide whether to adopt a tool, compare it with competitors, examine evaluation results and stop using it if it performs badly. They also bear, at least to some extent, the financial and reputational risks. A person whose asylum claim is being processed cannot choose the technological infrastructure of the institution deciding their case.
Another example of AI used in the asylum process, which pre-dates the generative-AI boom, is one of the projects developed by the Federal Office for Migration and Refugees (Bundesamt für Migration und Flüchlinge, BAMF). Since 2017, BAMF has used an automated language and dialect recognition system (DIAS) in asylum proceedings. As of December 2025, DIAS was still in operation.
AlgorithmWatch has documented criticism from linguists of the core assumption of this system: that a dialect, even if determined correctly (which is a big if), can be mapped precisely onto geographic origin. People move, languages mix and an individual’s speech can reflect family history or years spent somewhere other than their place of birth.
Soon after DIAS was introduced, VICE reported the case of an asylum seeker from the Kurdistan region of Iraq whose speech was classified as Turkish with 63 per cent probability. He told the publication that he did not speak Turkish. His language was Sorani Kurdish, which the system did not support, yet the automated analysis had still been carried out.
Despite DIAS having been used in asylum procedures for almost 10 years, a final independent evaluation report has still not been made public. BAMF commissioned the Nuremberg Institute of Technology to evaluate the system, and some very limited results have been presented publicly, but not the promised final report.
On 29 July 2026, the federal cabinet approved a draft law on the use of AI in migration administration. The proposal is to create broader legal grounds for authorities to use personal data from asylum and residence procedures to develop AI systems and other applications. The government says such systems could analyse administrative procedures and decision patterns, shortening processing times and increasing consistently. In certain circumstances, AI could also be used to compare information submitted in an application with publicly available information online.
The government explicitly states that individual assessments and decisions would remain the responsibility of human officials. It adds that administrative delays also have consequences for applicants as well.
PRO ASYL argues that the draft does not define clearly which authorities could use which AI systems and for what purposes, and raises concerns about transparency, personal data, discrimination and the ability of affected people to challenge algorithmically influenced processes.
Using AI for a professional is a choice. We decide which information to put into the chatbot, we can reject the answer, we can try another model, or decide to not use AI at all. This ability to choose is so ordinary that we barely register it as a form of power.
An asylum seeker is not choosing the chatbot into which someone else is entering their data, a person whose legal status is at stake cannot take their case to an alternative model provider.
The emerging AI divide is not only about access, it is not about who can afford the model or who is better at prompting. The divide is between the people for whom automation is an optional tool to make their life easier and the people for whom those tools become another obstacle to overcome, sometimes without even being aware of the existence of them.
What complicates the picture further is that the same people who are subjected to AI scrutiny might also use AI themselves, even to prepare applications for the very institutions assessing them. But the role we take when we become the object of an automated system, reduced to data and without any meaningful choice over whether the system is used, deserves much more attention.
Perhaps we, as a society, will not be able to stop the widespread AI adoption by government agencies; it promises too much convenience for the decision-makers. What we, however, should definitely demand is transparency and public oversight. The evaluation results should be made public before the system is applied to the cases of people whose lives might be severely affected by its decisions. The situation with DIAS, where after almost a decade of operation the full evaluation report is still not publicly available, should not become the norm.
And what about giving applicants the choice? They could be shown a description of the system, its known limitations and reliability, and told whether using it is expected to speed up the process and by how much. Of course the privilege of being familiar with the technology remains here. But at the very least, it would give people some agency over whether they want to be subjected to such systems at all.
