AMS (Arbeitsmarktservice Austria)
Austria's job agency AMS launches a ChatGPT career bot and gets gender-bias criticism
In January 2024 Austria's public employment service AMS launched Berufsinfomat, a career-advice chatbot based on ChatGPT, at a reported cost of about 300,000 euros. Within days testers showed it steering men and women toward stereotyped jobs, and AMS said it would keep improving the bot.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
about 300,000 euros
Reported cost of the Berufsinfomat project1
What it took
Cost, team size and timeline were not disclosed in the sources.
The problem
Austria's public employment service, AMS, wanted to give young people and jobseekers a faster way to explore career paths. It also hoped to relieve its human advisers.
What they built
AMS presented Berufsinfomat, a chatbot that is built on ChatGPT and linked to AMS's own occupation encyclopedia. It answers questions about careers and cites the AMS entries with links. An outside AI company that sells ready-made bots built it. Futurezone names that company as Goodguys.ai, and says the software was adapted only lightly from a standard product.
How they did it
The bot combines ChatGPT's general knowledge with AMS occupation content. AMS's chief Johannes Kopf said it was tested for months with advisers and users before launch. A hidden instruction told the model not to distinguish between men and women. AMS states a cost of about 300,000 euros, and shows a disclaimer that error-free answers are not guaranteed.
Results
The launch was overshadowed by testing from the public. Netzpolitik reports that for a man with excellent school grades the bot suggested IT and retail, and for a woman it suggested gender studies or philosophy. Futurezone reports similar patterns, with women pointed to care, beauty and fashion and men to software and warehouse work. A tester also got invented answers for a made-up job. AMS said it had trained the system to make as little distinction as possible and that work continued. Outside developers also reported code and security flaws.
What went wrong
The bias appeared quickly, in the hands of ordinary users. Kopf's posts, as read by netzpolitik, suggested AMS knowingly accepted such weaknesses in return for learning. A mathematician quoted by netzpolitik said the problem cannot easily be corrected after the fact. Privacy was a second issue: user inputs go to OpenAI, and AMS put the responsibility on users with warnings.
Lessons for your project
Test for bias with your own scenarios before launch. Outside testers found gender-stereotyped advice within minutes, even after months of internal testing.
A hidden instruction telling the model to treat men and women alike did not prevent biased answers. Measure outputs, do not rely on the instruction.
If you accept known weaknesses at launch, say so and set a date and a metric for fixing them. AMS's public stance that weaknesses are part of learning drew ridicule.
Check where user input goes. Critics noted that what users typed went to OpenAI, and AMS relied on warnings to users not to enter personal data.
Caveats
- The cost and the findings come from news reports. AMS gave the cost figure; the bias findings come from outside testers and journalists.
- No source read for this entry says whether the bot is still live or how it performed after January 2024. Usage numbers were not found.
- The AMS press release (3 January 2024) says the bot has been in use since early January 2024 and is open to all visitors of the AMS website; no exact go-live day is given, so the month is used.
- A second report (Der Standard) was not readable and is not cited.
Sources
- Diskriminierung: AMS erntet Hohn mit neuem KI-Chatbot
netzpolitik.org · 2024-01-05 · News · netzpolitik.org
- Diese sexistischen Antworten liefert der KI-Chatbot des AMS
futurezone.at · 2024-01-04 · News · futurezone.at
- Künstliche Intelligenz unterstützt bei Arbeitssuche
AMS Österreich (via APA-OTS) · 2024-01-03 · Press release · ots.at
Summarized in our own words from the sources above. Spotted an error? See corrections.