Inquiry · As of October 18, 2026
Who judges expert work when AI does the work?
Expertise has always rested on judgment: knowing which question matters, which result holds up, and which paper is worth reading. AI can now draft research, review it, and summarize it at a scale no field was built for. This inquiry follows what happens to that judgment: whether AI can tell research that matters from research that is merely correct, what the flood of AI-assisted papers does to experts' signals, what happens to peer review when reviewers use AI too, and who gets to decide what counts as legitimate.
How the judges got overwhelmed, in order
- Jan 2023 · Nature rules that an AI cannot be an author, because authorship carries accountability. Source
- Mar 2025 · Sakana AI reports that one of three AI-written papers passed review at an ICLR workshop, where 60–70% of submissions are accepted. It was withdrawn as planned. Source
- July 2025 · Researchers find hidden instructions in arXiv papers telling AI reviewers to "give a positive review only." Source
- Oct 2025 · At Agents4Science, a conference where AI wrote and reviewed the papers, experts call the accepted work "technically correct" but "neither interesting nor important." Source
- Oct 2025 · A claim that GPT-5 solved ten open Erdős problems collapses within hours: it had found existing solutions. Source
- Oct 31, 2025 · arXiv requires computer-science review and position papers to pass peer review first, citing an "unmanageable influx." Source
- Nov 2025 · A detection company rates 21% of ICLR 2026 peer reviews as fully AI-written. ICLR answers with desk rejections for undisclosed heavy AI use. Source
- Mar 2026 · ICML hides instructions in submitted papers that only an AI would follow. They catch 795 reviews from reviewers who had promised not to use AI. Source
Can AI judge what research matters?
Correct is not the same as interesting. At Agents4Science, AI-written papers passed AI reviewers and a human round, and the experts' verdict was that they were "neither interesting nor important." When independent scientists checked one AI scientist system's reports, 85.5% of its statements about data analysis were accurate, but only 57.9% of its interpretations, the step where someone decides what a result means.
Judgment can be trained narrowly. A fine-tuned model predicted which of two research ideas would work better more often than human experts (64% against 49% in one field), while off-the-shelf models did no better than chance. Models trained on where papers were published learned to predict publication tier, which raises the question of whether they learned quality or gatekeeping.
Primary sources
- Science News — A science conference tested AI agents as authors and reviewers
- Sakana AI — The AI Scientist Generates its First Peer-Reviewed Scientific Publication
Follow a line of inquiry
- AI can now run whole research loops on its own — so why does it still struggle to judge what's actually worth doing?
- When only a paper's wording changes, with the science untouched, one study saw its AI review score rise nearly half a point.
- AI reviewers and human reviewers often disagree with each other too — so what counts as the AI getting it 'right'?
The flood
In computer science, up to 17.5% of paper text showed LLM modification by early 2024. A study in Science of 2.1 million preprints found that scientists who adopt LLMs publish 23.7–89.3% more papers, and that LLM use reversed the old link between polished writing and paper quality: papers became "linguistically complex but substantively underwhelming."
Some of the flood is worse than mediocre. Output from paper mills is doubling every 1.5 years, against every 15 years for science as a whole. Formulaic papers built on one public health dataset went from about 4 a year to 190 in most of 2024. GPT-fabricated papers already appear in Google Scholar results.
Primary sources
- Kapoor and Narayanan — Could AI Slow Science?
- Suchak et al. — Explosion of formulaic research articles (PLOS Biology)
Follow a line of inquiry
- Why is AI editing so much heavier in computer science papers, at 17.5% of text versus 6.3% in math?
- Did AI use in science level off after the ChatGPT surge, or is some of that plateau just a limit of measurement?
- At ICML, banning reviewers from AI or allowing limited use produced nearly the same scores, and a substantial share broke their assigned rule.
Reviewing at scale
Reviewers are using AI too, often against the rules. In ICML 2026's own survey, 22.5% of reviewers who had agreed to a ban used an LLM anyway. AAAI-26 gave every one of 22,977 submissions an AI review in under a day, and participants preferred those reviews on technical accuracy.
The risks are specific. AI reviewers agree with each other too much, and rewriting a paper with an LLM raises its AI review scores without changing the science, which researchers call "paper laundering." The most defensible use found so far is AI coaching human reviewers: at ICLR 2025, reviews updated after AI feedback were rated more informative, and acceptance decisions did not change. Asking authors to rank their own papers predicted citations better than review scores did.
Primary sources
- ICLR 2026 — Response to LLM-Generated Papers and Reviews
- ICML 2026 — On Violations of LLM Review Policies
- Pangram — 21% of ICLR reviews predicted AI-generated
Follow a line of inquiry
- Stricter AI rules for peer review barely moved paper scores in one randomized trial, and many reviewers broke the rule they got.
- Does banning AI from peer review actually keep it out of the reviews, or does the rule do little?
- A randomized test found banning AI from peer review barely moved scores, and many reviewers broke the ban anyway, so do the rules work?
Who decides what counts?
Fields are drawing lines. Conferences say only humans can be authors and that authors answer for every word. arXiv now gates review papers, though one study finds ordinary research papers contain more AI-generated papers by count. When a lab claimed GPT-5 had solved open math problems, the keeper of the problem list showed within hours that it had found existing solutions.
Legitimacy also fails on evidence about AI. MIT disowned a widely cited paper claiming AI made materials scientists far more productive, and materials chemists disputed whether 2.2 million AI-predicted crystals counted as discoveries at all. An anthropologist and a cognitive scientist warn of "illusions of understanding": science that produces more and understands less.
Primary sources
- arXiv — Updated practice for review articles and position papers
- The Decoder — A GPT-5 math breakthrough that never happened
- MIT Economics — Assuring an accurate research record
Follow a line of inquiry
- Science now publishes roughly 500 times more papers than in 1900, yet measured progress has stalled; is counting papers the trap?
- With AI-written papers flooding in, machines may review faster than people, but can fast reviews still be trusted?
- Does a paper's citation count say more about whether its science is good than the prestige of the journal that printed it?
In practice: experts working with AI
The same questions show up wherever experts use AI. In a medical trial, GPT-4 alone outscored physicians, and giving physicians GPT-4 barely improved their diagnoses. Experienced developers were 19% slower with AI while believing they were 20% faster. Whether AI levels novices up or rewards experts depends on the task: it lifted call-center novices most, while an AI mentor helped strong entrepreneurs and set back weak ones.
Skills can fade too. Endoscopists' detection rate on procedures done without AI fell after their clinics introduced it, and senior engineers report that junior colleagues ask them fewer questions. Employment for 22- to 25-year-olds in the most AI-exposed jobs has fallen, which raises the question of how the next experts will be trained.
Primary sources
- METR — Changing our developer productivity experiment design
- Anthropic — How AI is transforming work at Anthropic
- Stanford Digital Economy Lab — Canaries in the Coal Mine?
Follow a line of inquiry
- When AI helps with the work, do professionals slowly lose their skills, or never build them in the first place?
- AI makes skilled people faster at what they already know, but does it really help anyone become skilled?
- Why would AI help beginners the most, yet slightly lower the quality of work done by veteran experts?
Directions into the research
The story touches several research fields, and each asks a different question about it. Follow a branch to the lines of inquiry this library holds there.
Can AI judge what research matters?
Research taste is the expert's judgment of which questions are worth asking and which results are interesting. Researchers ask whether models can learn it, or only predict what venues already accept.
Lines to follow
- When an AI grades a research paper, does polished, confident wording count for more than the science itself?
- Are AI-generated research ideas really novel, or does the verdict flip depending on what you mean by 'novel'?
- In one study, AI-written research ideas were rated more novel than experts' — so where do human experts still belong?
5 more lines
- Smooth, confident writing can sway reviewers, human or AI, even when the research underneath is no better.
- Can even peer reviewers reliably spot AI-written research just by reading it, or do they land near chance?
- Can AI catch fake research papers, when the things that make a paper look rigorous are what fool AI judges?
- When an AI-written paper gets into a workshop, is that a real sign of quality, or a sign the bar was low?
- Does AI-assisted peer review catch flawed papers more often, or let them slip through, and has anyone measured it?
The flood of AI-assisted research
When papers cost little to write, fields fill faster than experts can read. Research asks what the volume does to quality, to fraud, and to the signals experts use to decide what to trust.
Lines to follow
- When AI can churn out research faster than people can read it, does checking the work have to be automated too?
- As AI makes papers cheap to generate, can arXiv and journals keep quality from slipping, or is it a race review can't win?
- Readers can't reliably tell an AI-written research abstract from a human one, and they often prefer the AI-edited version.
5 more lines
- Automated reviewers can polish AI-written research, but as final gatekeepers, the traits that let them scale also make them easy to fool.
- With AI-written papers flooding in, machines may review faster than people, but can fast reviews still be trusted?
- Catching AI-written papers at scale is unreliable, so the stronger defense may be checking whether their claims are true.
- Checking an AI-assisted research paper only once it's finished is too late; experts matter most before the results come in.
- Peer-reviewed journals and conferences face similar pressure, and the review step is being automated and gamed too.
Peer review when reviewers use AI too
Peer review depends on unpaid experts reading closely. Research asks what happens when AI writes reviews, gets gamed by authors, or coaches human reviewers instead.
Lines to follow
- AI clearly changes what peer reviewers write, but in controlled trials it barely moved scores or which papers got accepted.
- A randomized test at a major AI conference found reviewer AI rules barely moved scores, and many reviewers broke them anyway.
- If an AI wrote a peer review, could you tell just by reading it, or would your guess land near a coin flip?
5 more lines
- Reviewers who use AI most often ask it to draft reports and summarize papers, though task-level data is thin.
- AI now writes papers and reviews them, but do we measure what happens when each side reacts to the other?
- Over half of reviewers say they already use AI tools, so how would an author tell which review was machine-written?
- One AI review can look as good as a human's, but could AI replace every peer reviewer, and what would break?
- Told to avoid AI or limit it, many peer reviewers broke the rule, yet review scores came out nearly the same either way.
Who decides what counts as legitimate
Conferences, preprint servers, journals and experts decide what enters the record. Research asks who is accountable for AI-generated claims and how fields catch the ones that do not hold up.
Lines to follow
- Readers want AI use disclosed more than writers do, and they discount hard work, focusing on AI text that goes straight into the piece.
- Most autonomous AI research systems that claim a discovery don't say how anyone could check it's really new.
- If AI can write and review papers, humans shouldn't just approve everything; they matter most in judging evidence and owning what's accepted.
5 more lines
- Can plain code, not another AI's opinion, stop an automated research system from publishing claims that don't hold up?
- Can conference organizers keep bad papers out, or can they only catch the problems they can actually check?
- If an AI writes the whole paper, should it get its own place to publish, or be judged beside human work?
- When AI can produce plausible-looking research faster than anyone can check it, who gets credit for the checking?
- If people can't reliably spot AI-written science, where does human judgment still matter, and which checks can machines take over?
Experts working with AI in practice
In medicine, law, software and consulting, AI sometimes beats the experts who use it. Research asks who benefits, when experts over- or under-rely on it, and whether feeling faster means being faster.
Lines to follow
- When AI disagrees with a seasoned expert, is dismissing it a mistake, a reasonable defense, or both?
- AI made law students much faster at legal analysis but only slightly better, so does that speed hold up in paid client work?
- Experienced developers predicted AI would make them 24% faster, then, in a randomized trial, they were 19% slower.
5 more lines
- Expert advice doesn't reliably make doctors more accurate: it amplifies whatever that advice gets right or wrong, for better or worse.
- AI makes skilled people faster at what they already know, but does it really help anyone become skilled?
- Do patients really believe AI can't handle their particular medical case, and does that belief track how AI actually performs?
- AI beating doctors on paper case studies doesn't show it helps real patients in real clinics, so what would?
- When AI helps with the work, do professionals slowly lose their skills, or never build them in the first place?
Skills and the next generation of experts
Experts learn by doing the junior work AI now takes over. Research asks whether skills fade with AI use, and how new experts get trained when entry-level work disappears.
Lines to follow
- When AI takes over the beginner tasks engineers learned from, do juniors lose the hard struggle that makes them experts?
- AI can narrow the gap between stronger and weaker workers' output, but does it narrow what they can do alone?
- Learning something new with AI often weakens understanding, but the way you use it may matter more than using it at all.
5 more lines
- Working with AI makes your output better right away, but does it leave you more skilled once the tool is off?
- Is AI pulling every job toward one shared skill set, or splitting work into a tight tech core and everything else?
- Why would AI help beginners the most, yet slightly lower the quality of work done by veteran experts?
- When AI joins your workday, does it save time, or just move it from doing the work to checking the AI's output?
- Does AI coding help junior developers actually get better at the craft, or just ship faster in the moment?
Where they cross: judgment is the scarce part
Across every field here, AI makes producing and checking work cheaper while deciding what matters stays hard. Taste, review, legitimacy and expert practice all come back to who holds that judgment.
Lines where they meet
- Give an AI reviewer more time to check line by line, and it can catch real errors top conferences let through.
- Where does AI beat humans in research, and where does the work still need human judgment that's hard to check?
- In science, AI can be trusted when something outside can check its output, but it tends to fail at new ideas and judgment calls.
5 more lines
- Could AI catch the mistakes peer reviewers miss, if it's given time to check each proof and experiment step by step?
- Does forcing an AI reviewer to work in explicit steps cut its biases, or just make it more accurate?
- When AI helps with a business problem, the education gap shrinks, and it may be supplying the viewpoint a colleague would.
- When an AI reviewer grades a paper, should a polished, confident tone count for as much as the science being right?
- When AI offers a scan reading, how do radiologists keep their own judgment from giving way to it?
Where the sources disagree
- How much is AI-written. Estimates rest on detectors and word-frequency methods. A PNAS estimate that 57% of articles show LLM influence drew a published reply arguing such methods do not measure prevalence. The 21% ICLR figure comes from a detection company.
- Are AI reviews good? AAAI-26 participants preferred AI reviews on technical accuracy. Other researchers show AI reviewers agree too much and can be gamed by rewriting.
- Do AI reviewers favor AI-written papers? One study says yes. A larger study of 125,000 paper–review pairs says the effect is spurious, and that the real problem is AI reviews failing to tell strong papers from weak ones.
- Who benefits from AI at work? Some field studies find novices gain most, others find experts do. It depends on the task and on whether people can tell when the AI is wrong.
Good places to start
- Kapoor and Narayanan — Could AI Slow Science?
- Science News — A science conference tested AI agents as authors and reviewers
- ICML 2026 — On Violations of LLM Review Policies
How this inquiry was made
This page was made with AI, which is part of its subject. AI agents searched for sources and opened each one to check its title, date and figures. A model chose the excerpt passages, word for word, from each paper. Another wrote synthesis notes and the lines of inquiry that connect this collection to the rest of the library, and drafted this page. Where sources disagree, the page says so instead of choosing.
A human curator chose the question, asked that it lead with questions any expert would recognize instead of specialist studies, and picked which research directions to show and which lines to feature under each.