Somewhere in the labyrinthine world of machine learning conferences, a quiet war is being waged. On one side: researchers who definitely wrote their peer reviews themselves, thank you very much. On the other side: organizers armed with invisible ink, hidden phrases, and the kind of paranoia usually reserved for spy thrillers.

The International Conference on Machine Learning (ICML) 2026 just rejected 497 papers - roughly 2% of all submissions - because the authors got caught using AI to write their peer reviews. And how did they catch them? By hiding secret instructions inside the PDFs that only an LLM would follow.

Yes, really. The organizers essentially honeypotted their own review process.

How to Catch a Lazy Reviewer

Here's the scheme, and it's beautifully devious: ICML embedded invisible prompts in every paper's PDF - text that humans can't see but that LLMs will dutifully process. These hidden instructions told any AI reading the document to include two specific randomly-selected phrases in its output. The conference created a dictionary of 170,000 possible phrases, then assigned each paper a unique pair. The probability of accidentally including the exact right two phrases? Less than one in ten billion.

When reviews came back containing the magic words, the jig was up. No amount of "I totally wrote this myself" could explain away a review that included both "synergistic blockchain metrics" and "parsimonious gradient descent" exactly as instructed.

The research behind this approach, published by Rao, Kumar, Lakkaraju, and Shah in PLOS One, found that certain watermarking strategies work remarkably well. Fake citations, for instance, appear in 98.6% of LLM-generated reviews. Even paraphrasing the AI output through another model only drops the detection rate to 94%.

The Bigger Picture Is Worse

ICML's 497 rejections are just one battle in a much messier war. Over at ICLR 2026, analysis by Pangram Labs found that 21% of peer reviews were fully AI-generated. That's roughly one in five reviews written by the academic equivalent of autocomplete.

The numbers tell a story of a system buckling under its own weight. ICLR submissions exploded from 7,304 in 2024 to 11,672 in 2025 to nearly 20,000 in 2026. NeurIPS, CVPR, and other venues face similar exponential growth. Meanwhile, the pool of qualified reviewers hasn't grown nearly as fast - senior researchers who traditionally review papers aren't multiplying like conference submissions.

So what happens when you need thousands of reviews and everyone's drowning? People cut corners. Over half of researchers in a 2025 Frontiers survey admitted to using AI while peer reviewing, despite most venues explicitly banning the practice.

The Irony Is Thick

There's something almost poetic about AI conferences struggling with AI misuse. These are the venues where researchers present work on making language models more capable, more helpful, more human-like - and then those same models get used to fake the very reviews that decide which papers get accepted.

ICML tried to address this by running two parallel review tracks: Policy A (no LLM use whatsoever) and Policy B (LLMs allowed for understanding papers and polishing reviews, but not generating them). Authors and reviewers could self-select. The 497 rejected papers came from reviewers who violated whichever policy they'd agreed to follow.

The watermark detection method isn't perfect, and ICML knows it. As they noted, this approach "may only catch some of the most egregious and careless uses" - reviewers who upload entire PDFs to ChatGPT and copy-paste the output directly. More sophisticated cheaters could evade detection by never feeding the PDF to an LLM in the first place.

What Happens Next?

The peer review system was already straining before language models entered the chat. Research output keeps climbing while qualified reviewer pools stagnate. Review times stretch longer. Quality suffers. And now there's an arms race between detection methods and evasion tactics.

Some researchers are proposing radical alternatives - federated conference models, author feedback on reviewers, formal accreditation systems. Others suggest embracing AI assistance transparently, with proper disclosure and human oversight. The one thing everyone agrees on? The status quo isn't sustainable.

For now, major conferences are experimenting with whatever they can. Hidden watermarks. Submission limits. Desk rejections for AI-generated content. It's whack-a-mole at scale, and the moles have GPUs.

The 497 researchers who got caught at ICML probably learned an expensive lesson about taking shortcuts. But the real lesson might be for the rest of us: the tools we build have a way of coming back around, for better or worse. And sometimes, the biggest threat to AI research conferences turns out to be AI itself.

References

1. Gibney, E. (2026). Major conference catches illicit AI use - and rejects hundreds of papers. Nature. https://doi.org/10.1038/d41586-026-00893-2

2. ICML. (2026). On Violations of LLM Review Policies. ICML Blog. https://blog.icml.cc/2026/03/18/on-violations-of-llm-review-policies/

3. Rao, V. S., Kumar, A., Lakkaraju, H., & Shah, N. B. (2025). Detecting LLM-Generated Peer Reviews. PLOS One. https://doi.org/10.1371/journal.pone.0331871 | arXiv:2503.15772

4. Pangram Labs. (2025). Pangram Predicts 21% of ICLR Reviews are AI-Generated. https://www.pangram.com/blog/pangram-predicts-21-of-iclr-reviews-are-ai-generated

5. Gibney, E. (2025). Major AI conference flooded with peer reviews written fully by AI. Nature. https://doi.org/10.1038/d41586-025-03506-6

6. Liu, X., et al. (2025). Position: The Current AI Conference Model is Unsustainable! arXiv:2508.04586. https://arxiv.org/abs/2508.04586

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.