No list of the "top" AI articles can be verified, because ranking depends on what a reader needs, but the type of document can always be checked. An arXiv preprint has passed a moderation screen and not peer review, a conference paper has been refereed (NeurIPS 2025 accepted 5,290 of 21,575 submissions, or 24.52 percent), and a news monitor such as the OECD's AI incidents tracker builds its summaries automatically. Knowing which of these is on the screen explains how much weight a claim can carry.
Key takeaways
- arXiv moderators check that a submission is appropriate and topical, and arXiv's submission page does not describe that screen as peer review.
- NeurIPS 2025 reviewed 21,575 valid main-track submissions with 20,518 reviewers; its chairs reported reviewer fatigue and noisy scores.
- Since October 31, 2025, arXiv's computer science category asks that review articles and position papers be accepted at a journal or conference first.
- The OECD AI Incidents and Hazards Monitor is a beta tool that collects news automatically, and its summaries are labeled as AI generated.
- A source's type says what checks it passed; it does not say whether the claim is right.
Why a "top articles" list is the wrong starting point
AI coverage comes in very different forms: research papers, company announcements, government guidance, newsletters, and automated feeds. A list that mixes them hides the differences that matter. A result reported in a paper that experts have reviewed is a different kind of evidence from the same result described in a launch post, and both differ from a headline that summarizes either one. Sorting by type first, and by topic second, tends to save more time than hunting for a ranked list that nobody can audit.
What each kind of source has been through
The table below uses statements published by the organizations themselves. It describes checks, not quality: a reviewed paper can still be wrong, and a preprint can be excellent.
| Source type | Check it has passed | Check it has not passed |
|---|---|---|
| arXiv preprint | A moderation process that verifies material is appropriate and topical, per arXiv's submission page | Peer review. The page asks for "refereeable" work but does not say arXiv referees it |
| Conference paper, such as the NeurIPS 2025 main track | Review by volunteers: 20,518 reviewers, 1,663 area chairs, and 199 senior area chairs handled 21,575 submissions | Independent replication. Acceptance records a judgment at one moment |
| Review article or position paper in arXiv computer science | Acceptance at a journal or conference, required before submission since October 31, 2025 | Workshop review, which arXiv says does not meet the standard |
| OECD AI Incidents and Hazards Monitor | Automated collection of news coverage, tagged by harm type, industry, and other fields | Human verification of each entry. The monitor is in beta and its content is not an official OECD view |
What the arXiv change shows about volume
On October 31, 2025, arXiv announced that its computer science category would stop accepting review articles and position papers unless they had already been accepted at a journal or conference, and that authors must supply documentation of that review. The announcement says these paper types were never an officially accepted content type and had been taken at moderators' discretion. It gives the reason in plain terms: submissions had surged, generative AI makes papers without new results quick to write, and most review articles received were "little more than annotated bibliographies." Volunteer moderators could not vet them, so arXiv chose to rely on refereed venues. The rule covers one category, and the post says other categories may adopt similar practices later.
For a reader the lesson is practical. Surveys of AI topics are now easier to produce than to check, and a survey posted to arXiv before that date carries no guarantee of review.
What peer review at scale looks like
The NeurIPS 2025 program chairs described a process under strain. They wrote that reviewers, area chairs, and senior area chairs "do not have infinite bandwidth," and they reported a new source of noise: reviewers raising scores to end author discussions while staying critical in private deliberations, which the chairs tied to reviewer fatigue. The 24.52 percent acceptance rate was, in their words, on par with previous years. The chairs also cited 11 confirmed cases in which papers co-authored by grossly negligent reviewers were desk-rejected. None of this means acceptance is meaningless. It means a score or a decision is a noisy signal, and that a single accepted paper is weaker evidence than several independent results pointing the same way.
Reading an automated feed with care
The OECD monitor shows how a useful resource can mislead if read as a record of verified events. Its page says collection is automated from news coverage, that one event may link to many articles, and that each entry's summary and its explanation for being labeled an incident or hazard are marked as AI generated. The monitor is labeled beta, and its content should not be reported as the official view of the OECD or its member countries. An entry is therefore a pointer to coverage that deserves a look at the underlying reports, not a finding.
Questions to ask of any AI article
- What type of document is it: preprint, reviewed paper, company post, regulator text, or news summary?
- Who wrote it, and does the author or organization sell the thing being described?
- Does it link to the data, code, or primary document, and does that document say the same thing?
- What is the date, and has the claim been revised, replicated, or withdrawn since?
- Are numbers given with their base, such as a count out of a total?
- Is the claim about a test result, a deployment, or a prediction?
FAQ
Is everything on arXiv unreviewed?
Not exactly. Many arXiv papers are later published in reviewed venues, and arXiv moderates submissions. The point is that moderation, as arXiv describes it, checks topic and appropriateness, so the posting itself is not a sign of peer review.
Does an acceptance rate of about one in four mean the accepted papers are the best?
It means they passed one process at one conference. The NeurIPS chairs themselves describe noise in scores, so acceptance is better read as a screen than as a ranking.
Can AI-generated summaries be trusted?
They can help with triage. The OECD monitor shows the safe pattern: it labels its summaries as AI generated, and a careful reader checks the linked coverage before repeating a claim.
Where should a newcomer start?
With primary documents when possible: a paper's own text, a regulator's guidance, or a standards body's publication. Secondary explainers are useful once the original is known.
Related reading
- What benefits of AI are documented
- A trillion-parameter model, examined
- Well-known AI technologies explained
- AI in self-driving cars
More in AI Basics.
Last verified: October 2026
Sources
- Attention Authors: Updated Practice for Review Articles and Position Papers in arXiv CS Category, arXiv blog, October 31, 2025
- Reflections on the 2025 Review Process from the Program Committee Chairs, NeurIPS blog, September 30, 2025
- Submission process and moderation, arXiv help pages, Cornell University, 2026
- AI Incidents and Hazards Monitor, OECD AI, 2026