Hidden prompts in papers submitted in 40th At the annual conference on Neural Information Processing Systems, if reviewers use an AI tool, certain words will be added to the report.
○Organizers of a prominent neuroscience conference are facing a backlash on social media after they added hidden prompts to papers to catch reviewers using generative artificial intelligence (AI) to review papers.
40th At the annual conference on Neural Information Processing Systems (NeurIPS), scheduled to be held in Sydney, Australia in December 2026, reviewers will be prohibited from uploading their peer-reviewed papers to an AI chatbot, as it violates confidentiality. Per the policy outlined in the conference handbook, reviewers can still use AI chatbots for background research purposes.
To enforce the policy and catch fraudulent AI use in peer review, event organizers intentionally hid instructions regarding large-scale language models (LLMs) in papers submitted for peer review.
This instruction instructs LLMs to use obvious phrases in their peer review reports, such as “this study addresses the central question” and “thesis argument.” Some researchers have already been caught trying to sneak secret messages into papers in an attempt to use AI tools to provide favorable reviewer reports. Many publishers prohibit the use of AI in peer review.
Several researchers reviewing the NeurIPS paper expressed concerns on social media about the indirect prompt injections inserted into the paper.
“Designing traps with malicious intentions in mind corrupts the relationships on which the entire system depends,” Søren Auer, a computer scientist at Leibniz University Hannover, wrote on LinkedIn. “You cannot build a healthy peer review culture if you treat reviewers as suspects.”
However, some see merit in this approach. A similar prompt injection effort found hundreds of reviewers abusing LLM in next week’s 43rd post.rd The International Conference on Machine Learning (ICML 2026) will be held in Seoul, South Korea, said Nihar Shah, a computer scientist at Carnegie Mellon University and the conference’s scientific integrity chair.
in a statement to transmitterthe NeurIPS organizing committee states that it cannot discuss the injected prompts in detail “without compromising the effectiveness of this intervention.”
ayou said transmitter He was assigned to review eight NeurIPS papers. He said he sometimes converts PDF files to Microsoft Word documents when conducting peer reviews, which provides some of the prompts.
Auer said he initially rejected the first paper he was reviewing because he thought the prompts had been inserted by the study’s authors. But after discovering a hidden prompt in the second paper and seeing researchers discussing the issue in a Reddit thread, he removed the flag.
You cannot build a healthy peer review culture by treating reviewers as suspects.
—Søren Auer
Many more papers may be rejected because reviewers don’t know the prompts were inserted by conference organizers, he says. “Personally, I don’t think it’s a good idea to ban the use of AI,” Auer added. “Of course, we need to discuss how to use it.”
The NeurIPS committee will respond directly to reviewers who notice the hidden prompts, informing them not to penalize individual papers, the statement said.
Like Auer, Sarah Atit, an AI researcher at the University of Surrey, said: transmitter She found the same prompts in all four papers she reviewed on NeurIPS. She says she also found it in her own version of the paper, which NeurIPS organizers prepared before sending it out for peer review.
Atit called the hidden prompt an “inadequate mechanism” and argued that while it may filter out some problematic submissions, it doesn’t solve the larger problem with peer review. “We blame reviewers too much because they are a visible point of failure,” she says.
But Shah said hidden prompts are “doable and achievable.” Shah led a similar initiative at ICML 2026, inserting hidden prompts in all submitted papers.
In doing so, Shah says he and his team have identified hundreds of auditors who used AI when they shouldn’t have, resulting in their reviews being rejected. ICML 2026 desk-rejected just under 500 papers for violating the LLM review policy. This represents approximately 2 percent of the total number of submissions received by the conference this year.
Researchers expressed “overwhelming support” for the strategy, Shah said, adding that they shared their methodology with the NeurIPS team. “I’ve been working on conference peer reviews for several years now, and I’ve rarely seen anything with such strong support,” he says. “People were really tired of reviewers copying and pasting AI-generated reviews without any effort.”
this article It was first published transmitter.
