ChatAmsterdam

Municipality of Amsterdam

ChatAmsterdam is a generative AI assistant developed for the City of Amsterdam to promote AI literacy within the organisation and to enable the safe and responsible use of AI as a tool in our work.

Last change on 2nd of July 2026, at 15:57 (CET) | Publication Standard 1.0

• Publication category: Other algorithms
• Impact assessment: DPIA, The Ethical Guide
• Status: In use

General information

Theme

Organisation and business operations

Begin date

2024-11

Contact information

[e-mail]

Responsible use

Goal and impact

Generative AI offers many ways to make work easier, such as finding information, summarising documents, analysing data or helping to write texts.  

 

However, this technology also has its downsides: training AI consumes a great deal of energy and natural resources; owners of the source material used to train the models are not compensated; and, depending on the sources, the model may produce racist or other ethically objectionable results. 

Apart from these general criticisms, there are also risks for the organisation. Documents submitted to a chatbot may be used by the supplier to further train the model. What if they contain sensitive information? Or what if a civil servant consults the bot to find information to support decision-making, whilst the model provides answers based on outdated or downright incorrect information? 

 

Research is needed to explore the value of Generative AI for Amsterdam whilst also identifying the practical risks. Amsterdam staff will therefore use an internal Generative AI chatbot, ChatAmsterdam, for a limited period. This chatbot has been built using OpenAI’s GPT-4o language model and has not been further modified or supplemented with (internal) training data. 

 

During the research period, usage patterns, value and risks will be assessed to determine whether we will use this technology in the future. 

Considerations

We know that Generative AI is already being used by staff, and we recognise both the opportunities and the risks. There are plenty of potential applications for the organisation, both internally and externally. However, before we authorise the use of Generative AI (within the compliance requirements) within the City of Amsterdam, further research is needed. How do people interact with a bot? What sort of (sensitive) information is included in the prompts? What are the results used for? How valuable is the technology for the City of Amsterdam and the city’s residents and businesses? 

These answers will help the City of Amsterdam to deploy the technology safely, responsibly and in a targeted manner. 

Human intervention

The chatbot and the underlying language model do not make independent decisions, nor are they part of an automated decision-making process that does not involve human intervention.  

 

All generated responses are reviewed by the user who asked the chatbot the question or gave the instruction. The user then acts on the response, for example by adopting it in part or in full, or by asking the chatbot further questions. The user is always the intermediary between the chatbot’s generated response and the task or process for which the response is being used.  

Risk management

ChatAmsterdam is offered as an internal chatbot to employees of the City of Amsterdam.  

 

The chatbot application, the data and the hosting environment comply with the applicable BBN2 security standards to minimise the risk of data breaches and security incidents.  

 

All prompts entered are stored and manually assessed for potential ethical risks, bias and any data breaches.  

 

In addition, there is a two-tier filtering system. Microsoft’s prompt filter screens all prompts for dangerous and unethical content such as discrimination, suicide or abuse. The second filter is an internally developed prompt filter that specifically searches for high-risk content such as the use of sensitive data, ethnic profiling, etc. 

 

These high-risk prompts are stored separately and submitted to the relevant Risk/AI officers for substantive assessment. 

Elaboration on impact assessments

Amsterdam is currently carrying out the ethical impact assessment in the form of an Ethical Information Sheet. This has been completed for this algorithm.

Impact assessment

• Data Protection Impact Assessment (DPIA)
• The Ethical Guide

Operations

Data

ChatAmsterdam uses the OpenAI LLM, specifically GPT-4o. This is the same language model (algorithm) that is also used in ChatGPT. 

 

This language model has not been further trained using additional (internal) data. The language model is used as it is, without any changes or fine-tuning. 

Technical design

ChatAmsterdam, as a generative AI chatbot, uses the GPT-4o language model in its current form, without the addition of any (internal) training data or other data. It is the same language model that ChatGPT uses. 

 

ChatAmsterdam works in a similar way to ChatGPT. Staff can ask a question or give a command (known as a ‘prompt’) to the chat application and then receive a generated response. An example of this is the summarisation or rewriting of texts. 

 

The language model processes the entered question or instruction and any accompanying information, and then sends the generated response back to the user. 

External provider

The algorithm (in this case, the GPT-4o language model) is provided by Microsoft via the Azure cloud platform.