ICML Rejects Papers of Reviewers Who Used LLMs
ICML Rejects Papers of Reviewers Who Used LLMs
The academic review system runs on a fragile assumption - that human experts read and evaluate papers. When reviewers outsource their reviews to LLMs, the entire system breaks. ICML's response highlights a detection problem with no clean solution.
The Two Detection Approaches
Prompt injection watermarks embed hidden instructions in papers that trigger predictable LLM responses. If a review contains the watermarked phrase, the reviewer used an LLM. This is clever but brittle - it only catches reviewers who paste the full paper into a model. Anyone who paraphrases or uses the LLM for parts of the review evades detection.
Statistical detection analyzes writing patterns - word frequency, sentence structure, vocabulary distribution. This catches more cases but has false positives. A non-native English speaker who writes in a style that happens to resemble LLM output gets flagged. A skilled prompt engineer who instructs the LLM to write naturally gets through.
Why This Is an Agent Problem
This is not just about academia. It is about any system that depends on verified human judgment:
- Code reviews - is the reviewer actually reading the diff?
- Legal reviews - is the lawyer actually analyzing the contract?
- Medical second opinions - is the doctor actually examining the case?
When agent-assisted work becomes indistinguishable from human work, systems built on the assumption of human effort need redesigning - not better detection.
The Real Fix
Instead of detecting whether an LLM was used, verify the quality of the output. A brilliant review is valuable whether a human or an LLM wrote it. A shallow review is worthless either way. Judge the work, not the tool.
Fazm is an open source macOS AI agent. Open source on GitHub.