Civinc

Municipality of Amsterdam

Using the Civinc application, where employees engage in anonymous conversations and sentiment analysis takes place using generative AI.

Last change on 13th of May 2025, at 11:18 (CET) | Publication Standard 1.0

• Publication category: High-Risk AI-system
• Impact assessment: Field not filled in.
• Status: In use

General information

Theme

Organisation and business operations

Begin date

Field not filled in.

Contact information

[e-mail]

Link to publication website

www.civinc.co/nl

Responsible use

Goal and impact

Civinc is a new application that allows employees to safely exchange views on important organisational issues with colleagues who have a different perspective . This is done in anonymous 1-to-1 chats. In demarcated sessions, employees with different perspectives can be engaged in large-scale discussions, without social desirability or peer pressure.

On the one hand, the application gives large groups of employees a voice and enables the organisation to discuss important (and sometimes sensitive) issues in a safe and inclusive way.

On the other hand, the application provides insight into employee sentiment, allowing the organisation to keep in touch with what is going on and thus avoid 'blind spots'. An AI analysis of the anonymous chats maps how employees view certain topics, which ideas emerge, and where the tension between different perspectives lies.

Considerations

This algorithm gives the organisation insight into employee sentiment on key organisational issues. Generative AI thereby enables large-scale textual data to be analysed and processed in a short time, in a way that would not be feasible/profitable with any other type of (manual) analysis. Thereby, the use of generative AI ensures that participants' chats do not need to be viewed, which is desirable from a privacy perspective.

The risk of bias or inaccuracies in the analysis is prevented by a quality assurance process that periodically checks whether the analyses performed are valid and reliable.

Human intervention

The output of the algorithm is always controlled by a human. In addition, the output of an analysis in this application may never stand alone, but must be considered and assessed in a broader context of data points and experiences.

Risk management

For the algorithm to meet high information security requirements, extensive compliance processes are in place.

Elaboration on impact assessments

A DPIA is not required because users share personal information only on a voluntary basis in anonymous chats or open text fields. Users are always advised against doing so.

A human rights check was carried out in the form of an Ethical Leaflet, Amsterdam model.

Operations

Data

The (generic) model is trained with data from OpenAI. This does not involve access to the data set used.

https://platform.openai.com/docs/models

Technical design

GPT-4(o) is a language model developed by OpenAI. It uses a transformer architecture, which is based on deep learning. The model is trained with large amounts of text data, learning to understand the structure of language. It can recognise context, make predictions about which words are likely to follow and generate meaningful answers based on the input text.

In the context of Civinc's application, specific instructions ('prompts') are given to the language model to provide relevant analysis to the aggregated anonymous chats in the application.

External provider

Civinc is the supplier of the application, which integrates OpenAI's algorithm (GPT4-o). More information on Civinc: www.civinc.co/nl/, more information on OpenAI: www.openai.com.