NWO bans secret AI prompts in grant applications

Research funder NWO has tightened its rules on AI. Notably, when applying for research funding, scientists are not allowed to ‘arm’ themselves against covert assessments by chatbots.

Many researchers regularly apply for grants from the Netherlands Organisation for Scientific Research. To determine which applications are awarded funding, the jury relies on the judgement of experts from the Netherlands and abroad.

Last year, NWO already ruled that these experts may not use generative AI when assessing grant applications. Uploading them to an AI programme would breach their confidentiality obligations, and there are also questions about the reliability of an AI judgement.

Hidden prompts

In an updated guideline, NWO has tightened the rules further. It now states that reviewers must formally confirm in the system that they have not used AI. Whether they actually refrain from doing so, however, is another matter. NWO receives thousands of applications each year, and reviewers – often researchers themselves – are busy.

The new guideline also states that applicants may not include ‘hidden prompts’ in their text. Last year, it emerged that scientists from countries including Japan and South Korea had embedded instructions specifically for AI bots in their papers, so that reviewers using AI would give a positive assessment.

These instructions appear in papers as white text on a white background or in a very small font, making them invisible to ‘ordinary’ readers. Tests showed that various chatbots were indeed more positive about papers containing such a hidden prompt.

Unnecessary?

At first glance, NWO’s tightening of the rules seems unnecessary, because if reviewers do not use AI, the hidden prompts have no effect whatsoever. However, an NWO spokesperson says it cannot be ruled out that an external reviewer may still use an AI tool.

And although AI is currently not permitted in the review process, ‘it could be a development that NWO may explore in the future’.

Approved tools

Outside the review process, NWO staff are allowed to use AI, but only if the tool has been approved. Last year, only the translation application DeepL was permitted. Microsoft Copilot Chat has now also been declared safe. Tools such as ChatGPT and Perplexity remain excluded.

According to NWO, AI detection software that could identify whether AI has been used at any stage of the process is not reliable and is therefore not used.