AI Papers, AI Reviews: Who Is Still Checking the Science?

AI-generated illustration of a researcher annotating a manuscript beside a laptop displaying AI feedback.
AI-generated illustration created for Routes, Bytes & Budgets: checking a research manuscript alongside AI feedback. This is a fictional scene, not a photograph of an ICLR review or a real software interface.


There is an uncomfortable loop developing in AI research: tools can help produce a paper, other tools can help review it, and somewhere in the middle a person still has to decide whether the science makes sense.

That last step is the one I keep coming back to. A convincing manuscript and a detailed review can look like evidence that serious checking has happened. But if both were generated with little human scrutiny, what exactly has been checked?

The debate around the International Conference on Learning Representations, or ICLR, makes this problem unusually visible. Submission limits, AI-use disclosures and penalties are becoming part of the response. The concern is real, although some of the dramatic numbers circulating around it need more careful labels.

First, what does “60,000 papers” actually mean?

AI researcher Gerard de Melo publicly reported more than 60,000 abstract submissions for ICLR 2027. That is a striking registration-stage figure. It should not be presented as an officially confirmed total of completed, valid papers entering review.

The distinction matters. ICLR’s author guidelines set separate deadlines: September 18, 2026 for abstracts and September 25 for full papers. They also specify that placeholder and duplicate abstracts are removed. A registration, a finished submission and a paper that survives initial checks are different things.

We do have a solid baseline. In its official retrospective for ICLR 2026, the conference reported:

  • 19,525 valid, format-compliant submissions.
  • 18,054 reviewers and 76,139 reviews.
  • 5,355 accepted papers.

That is already an enormous coordination task. And reviewers can assess several papers, so the number of reviews needed is not the same as the number of people needed. I don’t think the problem needs an inflated estimate to deserve attention.

The cost of checking has not fallen as fast as the cost of writing

A model can quickly produce an abstract, suggest an experiment or turn notes into polished prose. Establishing that an experiment answers a worthwhile question, or that an argument really proves its claim, is a different kind of work.

For me, that is the central imbalance. An author can save time generating material while passing a much larger verification burden to someone else. Even an incorrect criticism takes time to investigate and answer.

It would also be too convenient to blame the whole submission boom on chatbots. In their explanation of the new submission policies, ICLR’s organizers point out that rapid growth in machine-learning research predates today’s capable AI tools. They describe AI as adding pressure to a system already struggling to match research volume with expert attention.

More submissions alone cannot tell us how much useful research was produced, or how much of the increase AI caused. The harder question is whether the community can still give good work the attention it deserves.

The “21% AI-generated reviews” finding needs a caveat

In a November 2025 analysis of the ICLR 2026 cycle, AI-detection company Pangram classified 21% of the reviews it analyzed—15,899 reviews—as fully AI-generated. It classified more than half as having some level of AI involvement, including editing or assistance.

Those are the company’s detector-based estimates. They are not a verified count of misconduct cases. Editing a reviewer’s own assessment and generating an assessment from scratch also raise different questions.

ICLR’s 2026 organizers acknowledged the limits of detection. They used two detectors to flag reviews for area chairs to examine as part of assessing review quality. A flag was a reason for scrutiny, rather than an automatic verdict.

That seems essential to me. A system intended to protect research should not punish someone merely because their writing resembles a model’s output. The substance still needs examination: are the objections accurate, relevant and supported?

What ICLR’s new limits actually do

For ICLR 2027, the submission rules cap each author at 20 papers. There is also a narrower restriction: each author can appear on at most one submission in which none of the co-authors is an eligible reciprocal reviewer.

In practice, that second rule affects teams whose members lack the qualifying publication record defined by the conference. It is not a blanket rule allowing every newcomer only one paper, regardless of their collaborators.

Qualified authors appearing on three or more submissions must normally review at least six papers themselves, with exemptions for specified organizing roles.

I understand the appeal of connecting the demand for reviewing with a responsibility to contribute to it. But publication history is an imperfect proxy for expertise. The organizers themselves acknowledge possible disadvantages for interdisciplinary researchers and newcomers, as well as an incentive to add established names who have not earned authorship.

These are experiments in managing scarce attention. Their success should be judged by the quality and fairness of the reviews they produce.

AI assistance is allowed in some settings, with very specific boundaries

The rules differ across conferences and years. For ICLR 2027 reviewers, limited AI assistance with writing is permitted, but must be disclosed. Anyone using a model to generate or edit review text must also provide their original, self-written assessment and their interactions with the model. AI cannot be used to manufacture that initial assessment.

I like the question that requirement forces into the open: what did the reviewer actually think before the model rewrote it?

ICML 2026 took a different approach. Its Policy A prohibited LLM use in reviewing. Policy B allowed help with understanding the paper and polishing the review, using privacy-compliant tools, while prohibiting delegation of judgment and critique. Reviewers had to follow the policy assigned to them.

Enforcement was concrete. In March, ICML reported rejecting 497 submissions associated with 398 designated reciprocal reviewers who had violated their agreed LLM-use rules. The organizers said flagged instances were checked by humans and that generic AI-text detectors were not used for that enforcement.

Confidentiality belongs in this discussion too. An unpublished manuscript may contain someone else’s ideas, results or data. Whether a tool retains inputs or uses them for training depends on the service and its settings; conference permission is another requirement. This connects with the concern I raised about personal AI agents and control over our information: convenience should not quietly decide where sensitive material goes.

Could AI also help improve the process?

Yes, and this is where the story becomes more interesting than a simple ban.

ICLR partnered with Google to offer authors optional feedback through the Paper Assistant Tool during September 11–18, 2026. That feedback was private to authors and excluded from the official review process. The announcement also warned that the tool could miss real errors or incorrectly flag valid statements.

Helping an author spot a gap before submission could reduce work later. But the useful outcome is a corrected argument or a better experiment; accepting every suggestion would simply introduce another source of mistakes.

Meanwhile, NeurIPS 2026 is conducting a voluntary AI-assisted reviewing experiment. Participating reviewer–paper assignments are allocated to no assistance, open-ended assistance or structured assistance. Authors must opt in, and area chairs assess review quality without knowing the assigned condition. The conference says the participating model providers use zero data retention.

That is the sort of evidence I want to see: does assistance make reviews more accurate and helpful? Producing longer reports faster is a much easier target.

Someone still has to understand the result

My biggest concern is the possibility of agreement without independent checking. If an author and several reviewers lean on similar systems, a shared blind spot could pass through the process looking like consensus. That is a risk to investigate, rather than something we can measure from submission totals.

Human peer review has its own weaknesses. Still, a reviewer should be able to explain an objection, reconsider it when challenged and take responsibility for the judgment. A polished paragraph cannot do that on anyone’s behalf.

This brings me back to my earlier post about AI, mathematical research and human understanding. I want tools that help people understand more, catch mistakes and make stronger contributions. There is plenty of room for that.

But when producing a paper becomes easier, careful reading becomes more valuable. The real test of this new research workflow will be whether someone can still explain why the result is correct, why it matters and what remains uncertain.

Comments

Popular posts from this blog

Obsidian’s New Search: Less Friction for Notes in Two Languages

OPPO Find X9 Ultra vs Find X10 Pro Max: Is the Older Ultra Still Worth Buying?

OPPO Find X10 Technical Preview: Cameras, Design and Missing Specs