Please note: The algorithm descriptions in English have been automatically translated. Errors may have been introduced in this process. For the original descriptions, go to the Dutch version of the Algorithm Register.

DefGPT

DefGPT is the Ministry of Defence’s in-house alternative to well-known AI services such as ChatGPT. This service offers defence staff the opportunity to use Large Language Model (LLM) technology for a wide range of purposes, such as summarising or classifying texts, or as a tool for drafting new text or programming code.
Last change on 19th of June 2026, at 9:31 (CET) | Publication Standard 1.0
Publication category
Other algorithms
Impact assessment
DPIA
Status
In use

General information

Theme

Organisation and business operations

Begin date

2024-11

Contact information

https://www.defensie.nl/service/contact

Responsible use

Goal and impact

This algorithm has no direct impact on members of the public or businesses. The aim of the application is to help Ministry of Defence staff process and use information more quickly and securely. The tool supports day-to-day tasks, such as summarising texts, searching for information and drafting documents. In addition, the internal environment ensures that sensitive Ministry of Defence information remains protected and does not end up with public AI services.

Considerations

AI, and Large Language Models (LLMs) in particular, can make many tasks faster and more efficient. They assist with processing information, making decisions and utilising existing knowledge. This technology can therefore help the Ministry of Defence to prepare for its key tasks more quickly and effectively.


The introduction of DefGPT is a first step towards using this technology within the Ministry of Defence. It enables staff to work securely with an internal AI tool. This is important because sensitive Ministry of Defence information must not be shared with public services such as ChatGPT. This prevents others from gaining access to confidential information.

Human intervention

There is no automated decision-making involved. A human assesses the information generated by DefGPT.

Risk management

The chats are stored temporarily and then deleted. This data is not used to train the model.


In accordance with standards set out in, amongst others, the Baseline Information Security for Government (BIO), logging and authorisation take place within the application. This allows the use of the application to be monitored.

Elaboration on impact assessments

DefGPT does not process any of its own data as an application. The data entered by the end user is stored temporarily for the user’s convenience. It is then automatically deleted. Nor is the data used to further train the model within DefGPT.

Impact assessment

Data Protection Impact Assessment (DPIA)

Operations

Data

DefGPT uses OpenAI’s GPT-OSS 120B as its base LLM algorithm. OpenAI does not specify the specific sources on which the model was trained.

Links to data sources

OpenAI GPT OSS Model card: https://arxiv.org/abs/2508.10925

Technical design

LLMs (Large Language Models) are a type of artificial intelligence (AI) specifically designed to learn to mimic the structure and semantics of human language. They can generate text, translate, classify and answer questions based on a database of information. LLMs have applications in natural language processing (NLP) and are known for their ability to learn to understand complex linguistic structures and generate text that is difficult to distinguish from human language.


The textual input is broken down into a sequence of ‘tokens’ (words, letters or other text fragments) which are then fed into the model. The model then predicts what the next token is likely to be and uses this output to generate new text. This process is repeated until the desired length of the text is reached, resulting in a coherent and logical whole.

External provider

OpenAI

Link to code base

https://huggingface.co/openai/gpt-oss-120b

Similar algorithm descriptions

  • An application that helps staff at the Regional Self-Employment Department determine whether a business is viable.
    Last change on 5th of September 2024, at 13:06 (CET) | Publication Standard 1.0
    Publication category
    Impactful algorithms
    Impact assessment
    IAMA
    Status
    In use
  • To correctly determine eligibility of residents, we use the application of an algorithm as a source of information. For example, by pre-determining whether the required information fields are filled in the application.
    Last change on 31st of October 2025, at 13:39 (CET) | Publication Standard 1.0
    Publication category
    Impactful algorithms
    Impact assessment
    Uthiek, DPIA
    Status
    In use
  • Supports writers in formulating clear texts by assessing language level, marking difficult segments and suggesting improvements.
    Last change on 11th of March 2025, at 9:28 (CET) | Publication Standard 1.0
    Publication category
    Other algorithms
    Impact assessment
    Field not filled in.
    Status
    In use
  • This is a set of models that can assess texts. Based on the text in a registration, they determine whether it is about a particular topic of undermining. The outcome of the model is used to organise the work of police officers.
    Last change on 13th of May 2026, at 11:50 (CET) | Publication Standard 1.0
    Publication category
    Other algorithms
    Impact assessment
    Field not filled in.
    Status
    In use
  • Municipality of Almere, together with TU/e and Fontys, is investigating how generative AI can support the assessment of supplier documentation in tenders. The aim is to make the process more efficient without compromising transparency, accuracy and safety. The project ties in with ongoing TU/e and Fontys PhD research on trust in AI.
    Last change on 14th of October 2025, at 12:57 (CET) | Publication Standard 1.0
    Publication category
    Other algorithms
    Impact assessment
    Field not filled in.
    Status
    In development