One in Four Dutch Parliamentary Questions Now Written with AI Assistance
The Hague, Tuesday, 15 September 2026.
A recent investigation reveals that 26% of Dutch parliamentary questions are now AI-generated, raising critical concerns about politicians outsourcing their democratic thinking to automated algorithms.
The Rapid Growth of Automated Drafting in The Hague
A comprehensive investigation by the Dutch broadcaster NOS, analyzing 2,500 public documents containing more than 28,000 individual parliamentary questions, has revealed a significant shift in legislative workflows [1][3]. Over the past six months leading up to September 2026, 26% of written questions submitted to the Dutch government showed clear signs of artificial intelligence assistance [1][3]. This represents a substantial growth from the 17% recorded in the preceding six-month period [1][3], marking an increase of 52.941% in AI-assisted drafting. While the intensity of AI integration varies—ranging from minor stylistic polishing to entire blocks of text flagged as machine-generated—the trend highlights how deeply generative algorithms have integrated into the democratic processes of the Dutch House of Representatives [1][3].
How AI Innovation Works in Legislative Research
The practical application of generative AI in politics centers on accelerating research and refining communication. Members of Parliament, such as Daniël van den Berg of the JA21 party, utilize these systems to bypass hours of manual labor searching through extensive meeting minutes, parliamentary records, and historical media items [3]. Instead of replacing human input, the innovation acts as an intellectual accelerator: politicians use large language models like ChatGPT or Anthropic’s Claude to draft inquiries, summarize dense policy documents, and format complex queries more efficiently [1][3][4]. Additionally, representatives like Renate den Hollander of the VVD party utilize these tools for basic tasks such as shortening lengthy texts and performing spelling checks, while maintaining that the final political responsibility for the submitted content remains entirely theirs [3].
Balancing Efficiency with Democratic Oversight
Despite the clear efficiency gains, tech experts and political observers warn of substantial risks associated with outsourcing legislative tasks. Maarten Sukel, an AI developer, points out that while using AI is not inherently problematic, a critical risk arises when politicians rely on it to navigate complex policy areas in which they lack personal expertise [3]. In such cases, there is a danger that politicians might overlook factual errors or outsource their core analytical responsibilities to automated systems [1][3]. Furthermore, prominent tech thinker Marleen Stikker raises a systemic concern regarding digital sovereignty, questioning the wisdom of relying on proprietary, non-neutral algorithms owned by American AI companies [3]. However, the exact corporate headquarters and physical locations of these AI developers are not specified in the underlying investigative reports [alert! ‘source documents do not contain geographic location details for Anthropic or other AI creators’].
Broader Adoption Across the Dutch Cabinet
The integration of AI in Dutch governance extends far beyond written parliamentary questions. On Thursday, September 10, 2026, a report by de Volkskrant revealed that Prime Minister Rob Jetten of the D66 party used AI to write hundreds of social media posts [3][4][5]. Similarly, AI-generated passages have been identified in ministerial speeches [4]. An analysis using the AI detection tool Pangram revealed that 15% of a May 2026 speech delivered by Minister of Education Rianne Letschert at TU Delft was AI-generated, while 24% of an August 2026 speech by State Secretary Judith Tielen also showed traces of AI [4]. Although the Ministry of Education has defended the technology as a means to make government communications more accessible and understandable, they emphasize that ministers continue to define their own messaging [4].
The Challenge of Detection and Technical Safeguards
As AI usage proliferates, identifying machine-written text has become a technical battlefield. The NOS investigation relied on ‘Pangram’, an AI detector designed to recognize stylistic patterns from models like Claude, ChatGPT, and Gemini [4]. However, experts warn that these detectors are not foolproof; Pangram identifies less than 50% of fully AI-generated texts and can occasionally misidentify human writing as machine-generated if a person writes in a highly structured, formulaic style [3][4]. To counter this, AI developers are introducing native safeguards. For instance, Anthropic, the developer of Claude, has pioneered a statistical code that embeds an invisible watermark directly into its generated text, allowing detection tools to verify AI origin with high accuracy [4]. This comes ahead of planned European Union regulations that will eventually mandate digital watermarks for all AI-generated content [4].