AI is flooding ML conferences with papers faster than reviewers can scale

The opening post asks whether reviewers can keep up with more papers. Submission counts describe workload; they do not establish review quality or how much AI caused the increase. The accompanying image labels its 2027 figure as provisional identifiers, which cannot establish a final submission total.

ICLR, a machine-learning research conference, provides a documented example of the quality-control problem. Its organizers reported 76,139 reviews for the 2026 conference. They ran two AI-writing detectors over reviews and sent cases flagged by both to area chairs, who oversee reviewing. The flags informed human assessments of review quality. The organizers acknowledged that detection was imperfect and that AI-assisted writing could still contain sound judgments.

Checking references required human judgment too. An automated system compared citations with bibliographic databases and web searches, but sometimes flagged legitimate references, including translated titles. At least three people examined each flagged paper before rejection for confirmed fabricated references. The organizers’ account illustrates the distinction: automation identified possible problems, while people assessed the evidence behind decisions affecting authors.