2% of ICML papers desk rejected because the authors used LLM in their reviews

Original Article ↗ Hacker News Discussion ↗

A major machine learning conference quietly embedded invisible prompt-injection watermarks in submitted PDFs and found that around 2% of authors who had explicitly agreed not to use large language models still relied on them to generate peer reviews, leading to desk rejections. Commenters highlight that this figure is likely a lower bound, since the method only catches copy‑pasted LLM output and doesn’t cover reviewers who were allowed to use AI tools. The exchange explores broader concerns about honesty in academia, pressure on unpaid reviewers, the growing dependence on LLMs, and the security implications of models that cannot reliably distinguish content from instructions.

Detection method & scope

  • Many praise the hidden watermark / prompt-injection scheme in PDFs as clever, precise, and far more reliable than generic “AI detectors.”
  • It only flags reviewers who fed the full PDF to an LLM and pasted output verbatim, not those who used LLMs for light editing or idea support.
  • Several point out that the ~2% headline rate is very conservative; actual LLM use in both “no-LLM” (Policy A) and “LLM-allowed” (Policy B) groups is likely much higher.

Ethics, dishonesty, and dependency

  • Strong consensus that the core issue is not LLM use per se but breaking an explicitly chosen “no-LLM” commitment.
  • Some frame this as straightforward cheating and lying; others emphasize human weakness and “impulse control,” likening LLM reliance to addiction.
  • A few describe personal strategies (separate machines, blocking paste) to avoid LLM contamination of professional writing.

Debate over sanctions

  • Opinions range from “ban for life as a deterrent” to “this should be a learning moment, especially for students.”
  • Several argue research on deterrence suggests certainty of enforcement matters more than harshness of punishment.
  • Others stress punishment also signals community norms and rewards honest reviewers.
  • Clarified that reciprocal reviewers who violated Policy A had their own submissions desk-rejected; innocent authors are not targeted.

LLMs in reviewing: tool vs abuse

  • Some reviewers say they would (or do) use LLMs legitimately: summarizing, flagging issues, improving tone, or checking fairness.
  • Others insist that if you need an LLM to understand a paper, you shouldn’t review it.
  • There is skepticism that anti-LLM policies are sustainable amid rising workloads and paper volume.

Prompt injection & security concerns

  • Several note the irony that enforcement relies on the same prompt-injection vulnerability considered dangerous elsewhere.
  • The lack of separation between “data” and “instructions” in LLM inputs is highlighted as a fundamental security and reliability problem.
  • Commenters worry that authors can embed positive-review instructions in papers themselves, manipulating LLM-assisted reviewers.

Academic incentives & political economy

  • Multiple comments describe ML academia as hyper-competitive, low-trust, and overloaded, with reciprocal reviewing adding unpaid labor.
  • Some see LLM misuse as a predictable outcome of exploitative structures; others respond that reviewing is core professional service and conferences are not-for-profit.