Germany AI Asylum Law Draft Leans on Staff Training That Review Says Cannot Stop Automation Bias

Germany's proposed law on using AI in migration administration, called the KIMVG, relies on two protections against trusting computer results too much: people must review the results themselves, and staff who use the systems must receive special training. The bill does not use the term automation bias. But a 2010 review published in Human Factors found that this problem affects both beginners and experts and cannot be prevented through training or instructions. On Monday, 37 civil society groups, including AlgorithmWatch, published an appeal asking Germany's parliament, the Bundestag, to revise the bill before a public hearing by its Interior Committee. The hearing is scheduled for Nov. 23 at 8 a.m. ET (2 p.m. in Berlin), according to netzpolitik.
Anyone whose asylum or residence case is held by the Federal Office for Migration and Refugees (BAMF) or a local foreigners authority could be affected. The bill allows authorities to compare information about applicants with information from public internet sources. This means people who have the same name as an applicant, or who appear with that person in public material, could also show up in the results.
What the Bill Allows Authorities to Do With Case Files
The government's bill, Bundestag printed paper 21/8275, would change Germany's Asylum Act and Residence Act. One part would allow case data to be used to train, check and test AI systems. A second part would allow authorities to use AI to monitor how cases move through their offices. A third part would allow them to compare information about applicants with information from public internet sources.
The federal government says the aim is to make the process faster, more efficient and safer. The cabinet reviewed the main policy paper behind the plan on Nov. 5, 2025, and approved the bill on July 29, 2026. Germany's Interior Ministry counted 113,236 first-time asylum applications in 2025.
Human Review and Staff Training Are the Safeguards Named in the Bill
The bill says staff must manually check the results of internet searches and comparisons. Its explanation also says that automated comparisons cannot make decisions. It requires employees who use case-monitoring and internet-matching systems to receive special training. In its response to the Bundesrat, Germany's federal government said that the responsible staff would still examine each case and make each decision. Automated processes would not replace the legal review or make decisions that staff must follow.
Automation bias is a known problem. When people use a computer tool to help them make decisions, they often follow its suggestions, even when those suggestions are wrong. They may also fail to notice mistakes that the tool does not point out. In an essay on Verfassungsblog, constitutional law writer Arian Henning describes the human reviewer in these systems as a cover that makes the machine's role look less important than it really is. On the same website, authors Anna-Lena Priebe and Lise Känner argue that having a person make the final decision and requiring staff training are not enough to prevent automation bias. They also say that case monitoring and internet matching are likely to count as high-risk AI under the EU AI Act.
This concern is especially important because asylum decisions often depend on whether officials believe an applicant and how they assess the person's language. BAMF already uses software called DIAS to identify dialects. AlgorithmWatch describes this software as controversial. The bill's explanation gives one example of a reason to doubt an applicant's account: the person's spoken language or dialect does not match the region where they say they come from.
Anonymization Is Not a General Requirement Under the Bill
The bill does not create a general rule requiring personal information to be anonymized. According to the explanation of the bill, anonymizing or replacing identifying details with codes is not necessary if doing so is technically impossible, would cost too much compared with its benefits, or would seriously reduce the quality of the results. The explanation also says that names and dates of birth can provide useful context when checking whether information is believable. It mentions that real personal data might need to be used. The appeal from AlgorithmWatch and the other groups asks for anonymization to become the general rule.
Asylum case files can contain information about a person's ethnic background, religion and health. Article 9 of the GDPR treats these as special categories of personal data, which receive extra protection. According to an August report by netzpolitik, the office of the data protection commissioner, then led by Louisa Specht-Riemenschneider, wanted anonymized data to be used first, followed by pseudonymized data, with personal data used only when necessary. Moritz Hennemann has led the office since Oct. 1.
Does the EU AI Act Already Cover This? The Exemption and the Delay
In principle, yes. The EU AI Act lists AI systems that help assess asylum, visa and residence applications as high-risk systems under Annex III. Under Articles 6(3) and 6(4), a provider may argue that an Annex III system is not high-risk if it only finds patterns in decisions and does not replace or influence an earlier human assessment without proper human review. But this exception never applies to systems that create profiles of people. The provider must also document its claim and register the system. There is disagreement over whether a system trained on case files and used to flag cases would qualify. Priebe and Känner say the exception does not apply to the tools in the KIMVG.
The timing is also important. The Digital Omnibus, Regulation (EU) 2026/1744, took effect on July 27, 2026. It delayed the rules for Annex III high-risk systems until Dec. 2, 2027. This means those rules would not yet apply to a system put into use before that date.
Critics and Supporters Disagree About Whether the Safeguards Are Enough
The appeal from AlgorithmWatch and the other groups asks for anonymization to be the standard rule. It also calls for limits on the use of complete asylum files and data from central registers, stricter rules for internet searches and comparisons, and an assessment of the effects on basic rights that includes civil society groups. It also asks for greater transparency, ways for people to challenge decisions, and registration of the systems. Netzpolitik's report about the appeal mentions the Dutch childcare benefits scandal, described in Amnesty International's report "Xenophobic Machines". It also points to a UK visa process that used an algorithm and was withdrawn in 2020 after claims of discrimination. The UK Home Office disputed those claims when it dropped the tool.
Other people and groups have different views. According to netzpolitik, federal anti-discrimination commissioner Ferda Ataman raised serious concerns while the government was discussing the bill. The rights group Pro Asyl says the protections in the bill are not strong enough. Federal Integration Commissioner Natalie Pawlik supports the plan, saying AI could reduce waiting times.
The Bundestag held a first reading of the bill on Oct. 8. It lasted 20 minutes, after which the bill was sent to the Interior Committee. The Bundesrat asked the government to let internet comparisons also help establish a person's identity and nationality and work out possible return options. The federal government said it would consider this request. A 2023 ruling by Germany's Constitutional Court in Karlsruhe said that laws allowing automated data analysis must set clear and sufficient legal limits. However, that case concerned police laws in Hesse and Hamburg, not the administration of migration cases.
What to Watch Before and After the Hearing
The next important date is the hearing on Nov. 23. The bill could still be changed after it. According to netzpolitik, the invited experts include Alexander Steen of the German Informatics Society, Franziska Görlitz of the Society for Civil Rights and Eric Töpfer of the German Institute for Human Rights. There are three main things to watch: whether the bill deals with automation bias or adds protections beyond human review and staff training; whether anonymization becomes a legal requirement; and whether the final law promises to register the systems and carry out a fundamental rights impact assessment under Articles 27 and 49 of the AI Act. Applicants can ask authorities to show them the personal data they hold about them under Article 15 of the GDPR. This is the practical way to find out what information the authorities have.
Frequently Asked Questions
What is Germany's KIMVG?
The KI-Migrationsverwaltungsgesetz is a proposed government law. It would allow BAMF and foreigners authorities to train AI systems using case data, monitor how cases move through their offices, and compare applicants' information with information found on the public internet, according to printed paper 21/8275.
Can staff training prevent automation bias?
A widely cited 2010 review in Human Factors says that training and instructions cannot prevent automation bias. This is why critics question a bill that depends on staff training and manual checks.
Does the EU AI Act cover asylum decisions?
The EU AI Act lists AI systems that help assess asylum, visa and residence applications as high-risk systems under Annex III. However, the rules for these systems have been delayed until Dec. 2, 2027.
Does the bill require my data to be anonymized?
No. The bill does not have a general rule requiring anonymization. Its explanation says that anonymization or pseudonymization is not necessary when it is technically impossible, would be disproportionate, or would harm the quality of the results.
