evidoria

← Back to browse

Good practice

BAMF DIAS — Germany’s AI dialect identification system for asylum procedures

Germany · Nuremberg · See the Germany profile

BAMF’s DIAS analyses phonetic patterns in asylum applicants’ speech to verify claimed nationality. Deployed since 2017 with 15,052 analyses in 2021 and 85% accuracy for Arabic dialects; expanded to seven EU countries but criticised for a 15–20% error rate.

85 %
Accuracy — Arabic dialects (internally published)
73 %
Accuracy — Dari (internally published)
78 %
Accuracy — Pashto (internally published)
15,052 analyses
Analyses conducted, 2021 (2021)
9,923 analyses
Analyses conducted, 2020 (2020)
~7,800 analyses
Analyses conducted, first half of 2022 (H1 2022)
15-20 %
Critic-estimated error rate
BAMF DIAS — Germany’s AI dialect identification system for asylum procedures

Details

Maturity
Established
Promoter
Bundesamt für Migration und Flüchtlinge (BAMF)
Period
2017–present
Keywords
immigration, asylum, compliance, AI risk

Context

Germany's Federal Office for Migration and Refugees (BAMF) deployed the Dialect Identification Assistance System (DIAS) in 2017 to support caseworkers verifying asylum applicants' self-reported country or region of origin, by analysing phonetic features in a short speech sample and classifying the speaker's dialect into a geographic cluster.

Objectives

DIAS is intended as decision-support for human caseworkers, not an autonomous decision-maker, helping verify claimed nationality where other documentary evidence is limited.

Activities

By 2021 the system covered five Arabic dialect groups (Maghrebian, Levantine, Egyptian, Iraqi, Gulf), plus Dari, Persian/Farsi and Pashto. A subsequent European pilot extended the tool to Austria, Finland, Norway, Sweden, Greece, Switzerland and Lithuania.

Results

Internally published accuracy rates stand at approximately 85% for Arabic dialects, 73% for Dari and 78% for Pashto. Usage peaked at 15,052 analyses in 2021, up from 9,923 in 2020, before falling to around 7,800 in the first half of 2022.

Conclusions

The system has drawn sustained criticism from refugee advocacy groups, linguists and academic researchers, who argue dialect alone cannot reliably establish nationality given diaspora settings, register-shifting under stress, and overlapping border-region dialects. A 15-20% error rate is considered unacceptably high for outputs that can influence life-affecting asylum determinations, and information on how DIAS output is weighted in final decisions, and on how applicants can contest it, remains limited.

Implementation
Indicative cost
Medium (€50k–€500k) — No public budget figures disclosed; estimated as medium given multi-year national operation and international pilot extension across seven countries.
Time to results
Long (> 3 years) — Deployed since 2017 and still in active use through at least 2022, with an EU pilot extension — a long-running, multi-year deployment.
Staffing & skills
BAMF caseworkers use DIAS output as decision-support alongside other evidence

Conditions for success

  • Positioned explicitly as decision-support for human caseworkers rather than an autonomous decision-maker
  • Coverage expanded incrementally by dialect group (five Arabic clusters plus Dari, Farsi, Pashto) over several years
  • European pilot extension shared the tool across seven additional countries

Common failure modes

  • Independent critics report a 15-20% error rate, considered high for life-affecting asylum determinations
  • Dialect alone cannot reliably establish nationality: diaspora speakers, register-shifting under stress, and overlapping border-region dialects all undermine accuracy
  • Limited public information on how DIAS output is weighted in final determinations or how applicants can contest it

Where it fits

Governance type
national government agency, with EU-wide pilot extension
Scale
national, piloted in 7 additional European countries
Income level
high-income

Data sources

Where this practice's information was retrieved from, and when.

Attachments

Similar practices you may find useful