Germany's Draft Law on AI in Migration Raises Rights and Bias Concerns

In late July, Germany's cabinet quietly approved a draft law that would create a broad legal basis for using artificial intelligence in visa, residence, and asylum procedures.

The bill, known as the KI-Migrationsverwaltungsgesetz (KIMVG, or AI Migration Administration Act), would allow authorities to process personal data to develop, train and test AI systems, introduce AI-assisted procedural monitoring, and do an automated cross-checking of applicants' statements against publicly available internet sources.

The German government says the law is needed to address limited administrative resources and to improve the efficiency, quality, and security of migration procedures. Civil society groups, however, have identified legal gaps, what they describe as overly broad powers granted to authorities, and a long list of unanswered questions about how the systems would work in practice. In this article, we unpack those questions and highlight the need for a more inclusive deliberation of AI in migration and asylum administration.

What the bill would do

The German cabinet adopted the draft KIMVG on July 29. The proposed bill would amend the country's Asylum Act and Residence Act to allow AI use in three main ways. First, it introduces “automated procedural monitoring,” in which completed cases would be analyzed using AI to identify structural decision-making patterns and opportunities for standardization and faster processing.

Second, in cases involving “legitimate concern” that evidence is contradictory or implausible, it permits automated comparison of applicants' statements with internet sources, including social media.

Third, it would allow authorities to use personal data to train, validate, and test their own AI systems going forward. The bill leaves many technical details unspecified, however, so it remains unclear just how far AI use would extend —whether primarily to check information and assist with monitoring or also to contribute to decision-making itself.

The legislation, led by Germany's Ministry of the Interior, builds on a 2025 white paper from that ministry and Germany’s Foreign Office. After the first draft of the KIMVG was published on June 25, select civil society organizations were given roughly one week to submit comments, a timeline that every responding organization criticized. The German Caritas Association, for instance, said the short deadline made a thorough legal and technical review of the draft impossible. The timeframe of the bill's path through the Bundestag and Bundesrat, Germany's two houses of parliament, remains uncertain.

The stated rationale

Germany's government has consistently framed the case for the law around efficiency, new public management rationales and the country’s immigration interests. The 2025 white paper argued that automation would make procedures faster and less resource-intensive, reduce wait times, and benefit skilled-worker immigration in particular.

The cabinet draft points to steadily rising application volumes, constrained staffing, inconsistent administrative practice and legal interpretation across offices as justification for a “legally secure and data-protection-compliant framework” built around digital tools, particularly AI. This diagnosis captures a real problem: Germany's public sector does indeed face staffing shortages, and lengthy procedures burden applicants as well as frontline officials.

With respect to application numbers, however, the evidence is less clear, and in some instances the yearly numbers are actually declining. First time asylum requests, for instance, went down from 330.000 in 2023 to roughly 130.000 in 2025.

It is therefore remarkable that the draft presents AI as the only conceivable and supposedly cost-neutral solution, sidelining additional resources for immigration offices or simplified procedural rules as alternatives, as well as the costs introducing AI will entail. The rationale represents a broader tendency to solve what is essentially a political problem — staff shortages, resource allocation, complex administrative and legal situations — primarily through new technology. Research has cautioned against the assumption that introducing AI into public administration automatically produces cost savings and efficiency gains.

As the German Association for Data Protection (DVD) has noted in its statement, too, further digitalizing and introducing AI in Germany’s migration administration will first require additional staff, training and technical infrastructures – costs the draft does not account for.

Three lines of critique

While most of the civil society responses acknowledge that migration administration needs to be further digitalized and reformed, they also highlight three major concerns.

The first involves compatibility with data protection, constitutional, and EU law. Several organizations argue the scope of personal data use is too broad, and question whether sensitive personal data can be fully deleted once it has been used to develop AI systems. The advocacy group Pro Asyl argues the draft falls short of standards for fair administrative procedure and effective legal remedy, creating instead a basis for opaque, algorithmically driven decision-making. Likewise, the German outlet netzpolitik.org, which published its own analysis, noted that passages on automated data analysis closely resemble parts of a Hesse state policing law currently, which is under constitutional review.

Multiple organizations criticize that the draft does not fully meet requirements under the EU's General Data Protection Regulation (GDPR) or the EU AI Act. For instance, the draft bill invokes Article 10(3) of the AI Act, which requires training data to be “relevant, sufficiently representative, and to the best extent possible, free of errors and complete in view of the intended purpose”, mirroring GDPR's data accuracy standards.

As the German Association for Data Protection underscores in its statement, that is a demanding bar in the immigration context: the Central Register of Foreign Nationals (AZR), drawing on largely the same data sources the AI would probably be trained on, is known to have significant accuracy deficits, compounded by the fact that many of the relevant characteristics change quickly and go out of date.

The second concern is discrimination and bias. The draft requires authorities to avoid discriminatory algorithms without specifying how that should be ensured. Large language models (LLMs) generate outputs based on patterns learned from training data. If that data reflects existing biases in past migration decisions, those biases would also shape future predictions. Pro Asyl warns that, depending on the data selected, this context could indeed reflect or reinforce discriminatory stereotypes, while the German Association for Data Protection (DVD) has called for mandatory, independent external audits to cope with these challenges.

The third concern is the draft's openness and lack of specificity, particularly around how much decision-making authority AI systems would hold. The bill states that individual case decisions remain the responsibility of caseworkers, and according to the Ministry of the Interior, individual review stays “fully within the responsibility of the competent staff.” Critics counter that the draft does not explain how it would guard against automation bias — the tendency to trust machine-generated outputs over human judgment — meaning decisions could become de facto AI-driven in practice even if formally attributed to a human.

Questions left unanswered

Beyond the three main lines of criticism highlighted by civil society organizations, the draft leaves a long list of questions unresolved. What transparency measures would apply to migrants and asylum seekers whose cases are analyzed? How would caseworkers be trained to work alongside these systems? How would authorities prevent parallel, incompatible systems from emerging as digitalization proceeds unevenly across agencies? How would the systems account for the complexity of different procedure types and individual circumstances, including cases involving statelessness? And how, precisely, would the draft's protections for applicants' private lives constrain automated searches of their online activity?

The breadth of these open questions is itself part of the criticism: a law meant to establish a durable legal basis for AI in migration administration leaves many of the operational details — the parts that would determine whether the systems are fair, accountable, and effective in practice and how much AI would impact actual decision-making — to be worked out later, largely out of court.

That gap fits into a broader pattern. Governments and companies worldwide have faced criticism for rushing to apply AI, and generative AI in particular, to complex social and administrative problems faster than research, regulation, or public deliberation can keep pace. This dynamic is part of a broader “AI hype” in which the technology is treated as a default fix regardless of the problem at hand. Researchers and civil society groups have increasingly challenged the assumption that AI is necessarily an appropriate or the only tool for complex administrative problems in the first place. At the same time, governments are rapidly advancing the use of AI in migration administration, making its deployment increasingly likely.

This makes it all the more important to engage early with questions of how these systems should be designed and governed, who should be involved in shaping them, and what lessons can be learned from previous applications of AI in migration and asylum. One of the consulted civil society organizations, the Workers’ Welfare Association (AWO), has made that argument explicit, calling for an independent, interdisciplinary expert commission to examine — before any system is built — whether and under what conditions AI use in administrative procedures is compatible with rule-of-law requirements and the rights of those affected.

Such a more participatory approach is not only desirable from a democratic perspective but also matters for the success of digitalization processes. Research on past AI initiatives in migration policy lends weight to that caution, as can be seen, for instance, with Germany's heavily critiqued own dialect-recognition software, used to help verify asylum seekers' claimed origins, and the Netherlands' GeoMatch algorithm, which was halted during its pilot phase. Prior to that, research had shown how the GeoMatch algorithm optimizes in aggregate at the expense of individual outcomes, risking discrimination by ethnicity, gender, and marital status. Both examples suggest that top-down deployment without careful deliberation, piloting and input from affected communities carries rights risks for applicants and reputational risks for the governments involved.

Meaningful participation from civil society, advocacy organizations and migrants themselves could help shape not just oversight after deployment, but the questions that come first: whether AI systems are needed at all and what alternatives could be, what “quality” means in automated monitoring and what patterns are worth searching for.

Without that input, and without first settling whether AI belongs in these decisions at all, the KIMVG risks repeating a familiar pattern: treating a fundamentally political problem — staffing, resources, and complex law — as one that technology alone can solve, and asking those questions only after the legal groundwork has already been laid.