Summarising legal opinions on appeals

Municipality of Amsterdam

Summaries of existing appeals advice generated by generative AI (artificial intelligence). These support legal professionals in their need for information and enable a faster legal assessment of new appeals. 

Last change on 13th of July 2026, at 11:25 (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

2022-07

Contact information

[e-mail]

Responsible use

Goal and impact

Summaries are being drawn up of existing appeals advice. These summaries can be used to enhance the Case Library. This can be done by first displaying the key points of the case on the homepage. The Case Library is the internal database for the lawyers at the Legal Department (JB). The database contains advice on appeals and supports the lawyers in handling new appeals. 

 

The summary captures the essence of the appeal advice. This enables legal advisers to assess more quickly whether the appeal advice matches the information they are looking for. This solves the problem of having to spend a long time searching for the correct appeal advice, thereby saving time.

Considerations

This algorithm will ultimately enable legal professionals to consult appeals advice more quickly. As a result, processing times will be shorter and the assessment process more efficient.  

Human intervention

The results are available internally in the database of appeal recommendations (the Case Library). Anyone searching for appeals advice will be shown a summary of the full advice. The database includes a feedback function to report any errors in a summary, which are then corrected.  

Risk management

The data is only accessible within a secure environment. Legal professionals use it solely as a search tool. As mentioned, there is a feedback function to correct any errors in the data. It will also be stated that the summary was generated using generative AI. 

 

The results of the summaries will be checked on a random sample basis to prevent any incorrect output from the language model (based on the algorithm). This takes place during the fine-tuning process and will be prevented as far as possible through the use of prompt engineering techniques. 

 

There are no risks to vulnerable groups, exclusion or profiling  

Elaboration on impact assessments

Relevant elements from the IAMA have been incorporated into the Ethical Information Leaflet session

Impact assessment

• Data Protection Impact Assessment (DPIA)
• The Ethical Guide: https://open.amsterdam/woo-zoeken/detail/4ec7c01c-b112-4b54-9981-75d4e2827d1a/media/?mode=detail&view=list&rows=1&page=3&fq%5B%5D=search_s_dossiernaam:%22Ethische%20bijsluiters%22&sort=order_i_created_time%20desc

Operations

Data

Unstructured textual data from appeals decisions. 

Technical design

The algorithm has yet to be developed. Broadly speaking, data from appeal recommendations will be extracted from the dataset. Based on the guidelines we provide, the language model will generate a summary using the algorithm.  

 

OpenAI’s generative AI model, GPT-3.5-Turbo or GPT-4, will be used via Azure to summarise input texts in accordance with instruction-prompt techniques yet to be determined (for example, few-shot learning). This model can process textual input and generate new texts in natural language in accordance with the instructions provided by the user.  

External provider

The City of Amsterdam is developing this application itself.