BAMF dialect recognition

DIAS is dialect-recognition software run by Germany's federal asylum agency, the BAMF. It analyzes a short recording of an asylum applicant speaking and estimates the country or region their speech is most consistent with. Since 2017 the BAMF has used it to check stated origin when the applicant has no identity documents.

How accurate it is

The government's answers in the Bundestag, Germany's parliament, put recognition for Arabic at about 80 percent in 2017. So roughly one estimate in five was wrong. The figure was 87 percent in 2023, and about 75 percent for other languages in 2022.

Computational linguists judge some closely related varieties close to impossible to separate. Mark Liberman called separating vernacular Persian, Dari, and Pashto "probably pretty much hopeless." The varieties overlap, and a short sample cannot resolve them. The case file concludes that this is not a tool that is usually right with occasional errors. It is often uncertain by the nature of the task.

How it is used

The estimate is one input into a hard and important step: judging whether an applicant's stated origin is credible. That judgment can bear on whether they qualify for protection. Origin is often contested and hard to verify, so the case file calls the intent defensible.

In peer-reviewed fieldwork, BAMF decision-makers called the software "only a rough compass" and "too imprecise to solve the problematic cases." They treat its results as clues, not answers. The case file credits this framing as the responsible one. Used honestly as one clue among several, the tool is defensible.

What can go wrong

The documented risk is that an imprecise result gains more authority than its accuracy supports. Here the state's tool is set against the applicant's own account of who they are. A statement that the software finds their speech inconsistent with their stated origin is hard to rebut. It can carry more weight in the room and in the record than an error rate of about one in five, the government's 2017 figure for Arabic, warrants.

The BAMF also analyzes data from applicants' mobile phones. Fieldwork found this produced usable results in about a third of cases, and refuted a stated identity in about 2 percent.

Who can correct it

The applicant knows their own origin and has the most at stake. Yet applicants are often not shown the software's role or its estimate in a form they can contest. So an error they could explain at once may never surface. It could be a childhood across a border, an education in a second dialect, or a family that spoke differently from the region.

The stakes are what make this matter. As the case file puts it, a 20 percent error rate is one thing in a recommendation. It is another when it helps decide whether a person is returned to a country they fled.

How its use has grown

In July 2022 the BAMF extended the software to Farsi, Dari, and Pashto, and to more varieties of Arabic. It became the basis of a pilot by the EU's asylum agency, the EUAA, in seven countries. They are Austria, Finland, Norway, Sweden, Greece, Switzerland, and Lithuania.

This case starts with one pressure switched on, named Autonomy expands. Here it stands for that growth. On this network it adds to how much the software's estimates are written into the case record. It also adds to how far caseworkers rely on the software, and mistakes build on one another more.

Where the facts come from

The sources for this case are AlgorithmWatch's reporting from 2022 and peer-reviewed fieldwork by Scheel published in 2024. A third is a 2026 legal analysis by Beck in Verfassungsblog, a German law blog. The accuracy figures are the government's own, given in answers in the Bundestag.

What this network is drawn from

This network follows the pattern the case file documents. It does not reconstruct the actual system. It shows the dialect software, the phone data analysis, the applicant's speech sample, the caseworkers, the case record, and the reliability and challenge controls. The applicant is outside the network, and no asylum outcome is computed.

What the available tools can and cannot address

A failure pathway is a link between two parts of the network where a mistake made by one part can be passed on to the other. A tool closes a pathway when mistakes stop passing along it. The work along it may go on.

This case has a budget of 11 units. Understand the system, a Lab option that lowers some tools' prices, is not on offer here. Every price below is the full price. Explore (No Targets) sets no targets. The other three levels all ask for the network's mistakes to be contained, meaning corrected rather than building on each other.

Lingering effects are effects that stay after their cause is gone. The Lab's Dynamics menu starts at Both, which includes them, under Explore (No Targets) and Service Targets Only, where you can change it. Service and Safety Targets fixes it at Side-effects, which leaves them out. All Governance Targets fixes it at Both. With lingering effects on, and Understand the system not on offer, three tools work at reduced strength: Store less data, Review the riskiest first, and Check copied records.

Under Service Targets Only, the targets can be met, but no single tool meets them. The cheapest combination costs 4 units: Mark AI-written records with Escalate checks. Three more cost 5 units. They are Gate record entries with Escalate checks, Gate record entries with Mark AI-written records, and Mark AI-written records at its stronger setting with Escalate checks. With lingering effects off, Store less data with Escalate checks or with Mark AI-written records also meets them, for 5 units.

Under Service and Safety Targets, the case is not fully addressable with the available tools. This level also asks you to close every failure pathway. Eight are open before any tool is used. The first four are Origin estimate given to the caseworker, Caseworker weighs the clue, Assessment written to the case record, and Case record material analyzed. The other four are Caseworker reads the case record, Origin estimate entered in the case record, Speech sample analyzed for origin, and Phone findings given to the caseworker. No tool on offer acts on Caseworker weighs the clue, so that pathway stays open whatever you use.

Under All Governance Targets, the case is not fully addressable with the available tools either, for the same reason.

The sources describe no one running two checks on this network: Weight tied to measured accuracy, and Applicant challenge to the estimate. Check with a second model adds the first, and Assign a challenger adds the second. The challenger is a colleague the office names to question the big calls, not the applicant. No tool on offer lets the applicant challenge the estimate. Neither closes a failure pathway, and neither meets the targets at any level on its own.

More is not better here. Using every tool at its strongest setting costs 44 units, four times the budget. It contains the mistakes and closes every failure pathway but Caseworker weighs the clue. It still meets the targets at no level, because the software then contributes too little to the caseworkers' work.