Dutch Research Council Bans Hidden AI Instructions in Grant Applications
Utrecht, Saturday, 21 March 2026.
Scientists discovered manipulating AI reviewers through invisible text prompts embedded in research proposals face new restrictions from Netherlands’ top funding agency.
New Rules Target Deceptive AI Manipulation
The Dutch Research Council (NWO) has implemented revised guidelines in March 2026 that explicitly prohibit researchers from embedding hidden prompts in their grant applications [1][2]. These invisible instructions, designed to manipulate artificial intelligence systems used by reviewers, represent a sophisticated form of academic misconduct that has emerged in the digital age. The new policy requires assessors to formally confirm they have not used AI tools during the evaluation process, marking a significant escalation in NWO’s efforts to maintain integrity in research funding decisions [1][2].
The Hidden Prompt Problem Explained
The practice of embedding hidden AI prompts involves deliberately including instructions for chatbots within research papers using techniques that make them invisible to human readers [1][2]. Scientists from countries including Japan and South Korea have been discovered using white text on white backgrounds or extremely small fonts to hide these manipulative instructions [1][2]. When AI systems scan these documents, the hidden prompts can influence the assessment outcome, with tests demonstrating that chatbots become more inclined to provide positive reviews when such prompts are present [1]. This deceptive practice undermines the fundamental principle of fair peer review that has governed academic research funding for decades [GPT].
Timeline of NWO’s AI Policy Evolution
NWO’s current restrictions represent the culmination of a policy evolution that began in 2025 when the organization first prohibited experts from using generative AI to assess grant applications [1][2]. The initial ban was implemented due to concerns about confidentiality breaches and reliability issues when uploading sensitive research proposals to AI platforms [1]. By March 2026, NWO has expanded these restrictions to address the emerging threat of hidden prompts, creating a comprehensive framework that governs both reviewer behavior and applicant conduct [1][2]. The organization processes thousands of applications annually, making these safeguards essential for maintaining the credibility of the Dutch research funding ecosystem [1].
Approved AI Tools and Future Considerations
While maintaining strict controls over AI use in the assessment process, NWO permits its staff to use pre-approved artificial intelligence tools for other organizational tasks [1][2]. As of 2026, the approved list includes DeepL translation services, which gained approval in 2025, and Microsoft Copilot Chat, which has recently been declared safe for use [1][2]. However, popular AI platforms such as ChatGPT and Perplexity remain excluded from the approved tools list [1][2]. NWO acknowledges that AI integration in research assessment could be explored in the future, though no specific timeline has been established [1][2]. The organization has decided against implementing AI detection software, citing concerns about the reliability of such monitoring systems [1][2].