How Gemeente Amsterdam uses AI

The City of Amsterdam's legal department keeps an internal case library of its advice on objections (bezwaren) against municipal decisions. The city has registered a tool, built in house, that would use an OpenAI GPT model (gpt-3.5-turbo or gpt-4) through Azure to write a summary of the core of each existing advice, shown first in the library, so lawyers handling a new objection can judge more quickly whether an earlier advice is relevant.

Users can report errors in a summary, which are then corrected; the register says summaries will be marked as made with generative AI and checked by sampling. The register entry is marked "In gebruik" (in use) with a start date of July 2022, but its method section says the algorithm still has to be developed and that the prompting approach is yet to be decided, so the tool is recorded here as announced. No outcome figures are published.

The municipality of Amsterdam and its public health service (GGD) piloted AI avatar videos in the EU project CommuniCity to reach residents who do not speak Dutch or read with difficulty. Texts about vaccinations were translated with AI and spoken by a human like avatar in three languages for the GGD vaccination page. Two human checks applied: a communication adviser or translator reviewed the translated texts and the texts the avatar speaks. The register entry says early results were promising and the videos were watched a lot, without giving figures. The entry runs from May 2024 and is marked as out of use since February 2026.

Amsterdam's enforcement team received reports from neighbors and rental platforms about homes rented to tourists beyond the limits of 30 nights a year and four guests at a time. A random forest regression model, trained on five years of earlier illegal rental cases, residents' registry data and building data, estimated for each new report the likelihood that the address was being illegally let, so that limited enforcement capacity went to the likeliest cases first.

SHAP values showed officers which indicators drove each score, no automated decisions were taken, a person was appointed to monitor the model and staff received a work instruction and a workshop on the opportunities and risks of algorithms. The responsible inspector and the project enforcer determined actual violations through desk and field investigation. The register lists a start date of July 2022 and now shows the algorithm as out of use; the entry gives no reason and no performance figures.