The International Conference on Machine Learning (ICML) opens today at the COEX Convention Center in Seoul, South Korea, running from July 6–11. But the premier AI gathering — one of the fastest-growing in the world — arrives under the shadow of a landmark integrity crisis that has sent shockwaves through the global research community.
ICML 2026 organizers announced they desk-rejected 497 papers — roughly 2% of all submissions — after catching reviewers using large language models (LLMs) to generate peer reviews, in violation of policies they had explicitly agreed to.
The detection method was as ingenious as it was controversial: conference organizers embedded hidden watermark instructions into the PDFs of submitted papers. Invisible to human readers, these watermarks instructed any LLM that processed the PDF to include two specific, randomly selected phrases in its output. The phrases were drawn from a dictionary of 170,000 options, making the probability of accidental inclusion less than one in ten billion.
When flagged reviews were manually verified, 795 reviews from 506 unique reviewers were confirmed to have violated Policy A — the strict "no LLM" rule. Of those, 51 reviewers were so egregious — using AI in more than half their reviews — that all their reviews were deleted and they were permanently removed from the reviewer pool.
ICML 2026 offered reviewers two policies to choose from: Policy A (Conservative) strictly banned LLM use at any stage of reviewing, while Policy B (Permissive) allowed AI for understanding papers, literature searches, and grammar polishing — but still required human judgment for evaluation. Every reviewer assigned to Policy A had explicitly selected it or indicated they were okay with either policy.
"This is simply a statement that the reviewer used an LLM at some point when composing the review, which is unfortunately a violation of the policy they agreed to abide by," the ICML 2026 program chairs wrote in a blog post.
The scandal has broader implications. At ICLR 2026, independent analysis found that 21% of peer reviews were entirely AI-generated, with over 50% showing signs of AI use. The systemic nature of the problem suggests the academic honor system for peer review is breaking down under the pressure of hyper-competitive AI research.
ICML 2026 itself received approximately 25,000 submissions — a volume that reflects both the explosive growth of AI research and the strain it places on the peer review system. Critics argue the punishing workload on reviewers may be driving the cheating, while defenders of the crackdown say trust is the foundation of scientific progress.
The conference's keynote speakers, tutorials, and workshops beginning today in Seoul will cover topics ranging from reasoning limitations in current AI systems to quantum machine learning and trustworthy AI. IBM, Google DeepMind, Apple, and major universities worldwide have a strong presence at the event.
ICML organizers acknowledged their watermark method was "not difficult to circumvent, particularly if it is known publicly" and may only catch "the most egregious and careless uses of LLMs." The broader lesson, they said, is that "as our field changes rapidly the thing we must protect most actively is our trust in each other."
As the AI community gathers in Seoul this week, the 497 rejections serve as both a warning and a reckoning: if the architects of AI cannot be trusted to use their own creations responsibly, what message does that send to the world?
