AMS Berufsinfomat — Austria's Public Career-Information Chatbot and Its Bias Scandal

The Berufsinfomat is a free generative-AI chatbot launched on 3 January 2024 by Arbeitsmarktservice Österreich (AMS), the Austrian public employment service, following a proof of concept commissioned in April 2023. It was built by GoodGuys GmbH with Bundesrechenzentrum GmbH and Braintrust Marketing Service GmbH. The system answers questions about occupations, salaries, education and training pathways by retrieving from AMS's own content — the Berufsinformationssystem, Berufslexikon and AusbildungsKompass — and passing it to a commercial large language model via the OpenAI API; the ministry states that no AMS data were used to train a model. It is used by pupils and career changers directly, and by AMS counsellors and the BerufsInfoZentren, including in workshops with school classes.

Published usage figures are substantial. Roughly 140,000 questions were asked in the first two and a half weeks, to 20 January 2024. AMS reports 122,150 answers generated during 2025, in more than 100 languages. For scale, the same 2025 AMS report records 1,857,200 visitors to the Karrierekompass and 711,560 users of the Weiterbildungsdatenbank, so the chatbot remains a small channel next to AMS's established career-information products. Development costs billed to 31 December 2023 came to 275,160 euro — 74,880 euro to GoodGuys GmbH for the proof of concept, 14,280 euro to Braintrust and 186,000 euro to Bundesrechenzentrum — with a projected maximum of 231,977.51 euro for the first quarter of 2024. The contract was placed by direct award, without competing tenders.

Within 48 hours of launch the tool became a national controversy. Journalists and independent developers documented gender-stereotyped and discriminatory answers: women steered toward part-time and care occupations, gay men toward "creative" jobs, and HIV-positive users offered assistive-technology advice. Security and code weaknesses were also exposed publicly. Answering parliamentary question 17485/J, the responsible ministry conceded that at launch not all points could be implemented satisfactorily and that errors cannot be entirely excluded. Corrections shipped on 5, 10 and 19 January 2024. AMS had flagged around 5,000 of the first 140,000 questions as prone to gender bias, built a 1,000-question gender test set, and ran regression tests on samples of 100 to 200 random questions. The source code was also obfuscated after the criticism.

The documentary transparency here is unusually strong: costs, procurement route, bias-mitigation method and annual usage counts are all public, via parliamentary answer and AMS's annual reporting. Outcome evidence is absent. There is no impact evaluation, no controlled study, no published user-satisfaction data, and nothing linking use of the chatbot to better-informed education or occupational choices. Austria already carried a related precedent in the AMS jobseeker-profiling algorithm (AMAS), which makes a gender-bias failure in a successor AMS system a repeat rather than a surprise. The case is worth adopting as a model of disclosure and of post-launch remediation under public scrutiny — not as evidence that AI career guidance improves decisions.

Read the full analysis: https://www.ams.at/arbeitsuchende/aus-und-weiterbildung/berufsinformationen/berufsinformation/berufsinfomat