Summary: AI is turning weak incentives in academia into an industrial-scale problem: if careers, funding, and publisher revenues reward more papers, machines will produce more papers.

The core risk is not only fake research, but polished, plausible, low-value or misleading papers overwhelming editors, reviewers, grants, citations, and eventually AI systems trained on polluted literature.

The sector needs to shift from detection to provenance: stronger submission triage, verified data/code/workflows, stricter public-dataset rules, AI-use disclosure, shared integrity infrastructure, and less reliance on publication counts.

The Infinite Paper Machine: How AI Is Breaking Peer Review Before It Fixes Science

by ChatGPT-5.5

The Wired article “AI research papers are getting better, and it’s a big problem for scientists” describes a problem that is more serious than “students using ChatGPT” or “authors getting writing help.” It points to a structural crisis in scholarly communication: AI is making it cheap, fast, and increasingly difficult to detect the mass production of plausible research papers. The danger is not only that some papers are fake. The deeper problem is that the entire research system still rewards volume, citations, publication counts, and apparent productivity. AI has now entered that system as an industrial amplifier.

The result is predictable: if academia rewards papers, machines will produce papers. If careers depend on publication counts, CV padding becomes rational. If publishers earn revenue from article-processing charges, more submissions can look like growth. If peer review depends on unpaid academic labour, the cost of quality control is pushed onto already exhausted researchers. AI does not create these incentives, but it makes their failure visible at scale.

1. What the article says is happening

The article opens with a revealing example: a 2017 paper by Peter Degen and colleagues suddenly began receiving an unusual number of citations. Degen discovered that many of the citing papers followed a similar pattern. They used public datasets, especially large health datasets, to generate endless variations of statistically framed papers: disease X in population Y, risk factor A correlated with outcome B, association C among demographic group D.

The important point is that these papers were not always obviously fake. They were often formulaic, weak, misleading, or scientifically trivial, but no longer filled with the absurd errors that made earlier AI-generated papers easier to spot. The article contrasts older examples of AI sloppiness — hallucinated diagrams, odd phrases, fake citations — with the newer problem: AI-generated papers that are coherent, structured, adequately referenced, and just plausible enough to demand serious editorial attention.

That is where the crisis lies. Bad papers used to be easier to reject quickly. Now many of them are polished enough to consume expert time.

The article gives several examples of this shift:

A researcher at the University of Surrey noticed multiple similar submissions using the NHANES public health dataset. These papers scanned for correlations and then framed them as publishable findings, even where the relationship was likely meaningless or misleading.

Editors reported large increases in submissions. Security Dialogue saw submissions rise 100 percent over the prior year. Accountability in Research reportedly saw a 60 percent increase. Quantitative Science Studies saw a 40 percent increase. These increases matter because peer review does not scale like paper generation does.

Editors are also struggling to find reviewers. Where a few invitations once produced enough reviewers, editors now sometimes need to send 20 requests to secure two responses. That signals not only submission inflation, but reviewer exhaustion.

The article also highlights the next stage: agentic scientific tools that can analyse data, generate hypotheses, produce charts, write papers, and cite literature. One example described in the article produced a decent-looking paper from a dataset in under 26 minutes. This is not merely AI-assisted writing. It is the partial automation of the research-paper production pipeline.

2. The main issues

A. The collapse of the old fraud-detection model

The older model of research-integrity detection relied on spotting artefacts: plagiarism, duplicated images, tortured phrases, hallucinated citations, nonsensical diagrams, obvious statistical errors, and copy-pasted AI disclaimers. That approach works when fraudulent or low-quality papers look abnormal.

The article suggests that this window may be closing. Newer AI-generated papers increasingly look normal. They may cite real papers. They may contain plausible charts. They may use conventional structure. They may avoid obvious hallucinations. This creates a dangerous grey zone: the editor cannot easily tell whether the manuscript is fraudulent, AI-assisted, weak but legitimate, or written by an inexperienced researcher.

That is a much harder problem than plagiarism. Plagiarism asks: “Was this copied?” AI-generated research asks: “Was this work genuinely done, is it meaningful, and can its claims be trusted?”

B. The rise of “technically correct but scientifically useless” papers

One of the most important distinctions in the article is that not all of these papers are necessarily fake. Some may contain real statistical analysis. Some may use real datasets. Some may produce real correlations. But that does not mean they create knowledge.

This is a crucial point for scholarly publishers. Research integrity is not only about fraud. It is also about relevance, novelty, interpretation, methodological soundness, and contribution. A paper can be non-fraudulent and still be junk science. AI makes this category explode.

Public datasets are especially vulnerable. If anyone can take a large dataset and run thousands of pairwise associations, then the system will produce endless publishable-looking “findings.” Many will be random, trivial, context-free, or misleading. The scientific literature then becomes polluted with claims that are hard to disprove individually but collectively degrade the knowledge base.

This is the “spamification” of science.

C. The asymmetry between generation and verification

The article’s most important operational insight is the asymmetry between producing and checking papers. AI can generate a manuscript in minutes. A serious reviewer may need hours or days to assess methods, citations, data, novelty, and interpretation.

This asymmetry is fatal for a system built on voluntary expert review. The paper mill or AI-assisted author pays almost no marginal cost to generate another submission. The journal, editor, reviewer, and downstream scientific community absorb the verification cost.

That is not sustainable.

In cybersecurity terms, the attacker has automation; the defender has human labour. Unless the sector changes the economics, the defender loses.

D. Peer review is being asked to do too much

Peer review was never designed to be a forensic fraud-detection system at industrial scale. It was designed as a quality-control process among experts. The article shows that editors are now expected to detect fake citations, AI-generated text, manipulated datasets, statistical nonsense, paper-mill patterns, hallucinated literature, and possibly AI-generated peer reviews.

That is an impossible burden if the submission volume continues to grow.

The sector often romanticises peer review as a moral institution. In reality, it is a fragile labour arrangement dependent on unpaid or underpaid expert goodwill. AI exposes that fragility.

E. AI-generated peer review may compound the problem

The article notes that researchers are already using AI to assist with peer review, and that some studies suggest AI-generated reviews can miss methodological problems, cite retracted studies, and prefer AI-generated writing.

That creates a circularity risk: AI-generated papers reviewed by AI-generated reviews, processed by overburdened editors, and then cited by AI tools trained or grounded on the resulting literature. This could create a self-reinforcing knowledge-laundering loop.

For scholarly publishing, this is one of the most dangerous scenarios: the literature becomes not merely noisy, but recursively contaminated.

F. The incentive system is broken

The article rightly moves beyond technology and identifies the underlying incentive problem. Universities, funders, promotion committees, and some national systems reward publication counts, citations, and journal placement. Publishers may benefit commercially from higher submission volumes, especially under APC-driven open-access models. Researchers under career pressure have strong incentives to publish more.

AI is therefore entering an environment that already confuses quantity with quality.

This is the central lesson: AI is not an alien force attacking a healthy system. It is exploiting the weaknesses of a system that already rewarded excessive paper production.

3. Current consequences

The immediate consequences are already visible.

First, editors are overwhelmed. Submission volumes are rising sharply, and triage is becoming more labour-intensive because poor papers are harder to identify quickly.

Second, peer reviewers are becoming harder to recruit. If reviewers receive more requests while also facing pressure to publish more themselves, the system enters a fatigue spiral.

Third, journals risk publishing more weak, misleading, or fabricated work. Even if only a small percentage gets through, the absolute number could be large because the volume is growing.

Fourth, trust in the literature declines. Readers, clinicians, policymakers, journalists, and AI developers may find it harder to distinguish high-quality evidence from polished but empty output.

Fifth, citation systems become distorted. AI-generated papers can cite each other, cite convenient but non-canonical sources, or inflate citation counts for particular papers. That damages metrics that are already overused in assessment.

Sixth, public datasets become paper factories. Datasets created for legitimate research may be mined repeatedly for shallow correlations, creating the illusion of productivity while adding little knowledge.

Seventh, research integrity teams face an arms race. Detection tools can help, but the article suggests detection will become less reliable as AI improves.

4. Future consequences

The future consequences are more serious.

A. The literature may become polluted faster than it can be cleaned

Retractions are slow. Corrections are slow. Investigations are slow. AI-generated paper production is fast. If the rate of pollution exceeds the rate of correction, the scientific record becomes progressively harder to trust.

This matters especially for medicine, public health, climate, psychology, education, and policy-relevant fields where flawed findings can influence decisions.

B. AI systems may ingest corrupted science

Future AI tools for researchers, clinicians, regulators, and students will increasingly depend on scholarly literature. If that literature contains more AI-generated filler, weak correlations, fake citations, and low-quality synthetic papers, AI systems may reproduce and amplify those weaknesses.

This is a provenance crisis. Bad papers do not stay inside journals. They become training data, retrieval sources, citations, systematic-review inputs, policy evidence, and clinical-decision-support material.

C. The prestige economy may become even more unequal

Researchers with access to better AI tools, better data, better editing systems, and stronger institutional support may produce more papers and gain more citations. The article mentions research suggesting AI adopters publish more and receive more citations. That may advantage already well-positioned researchers while increasing pressure on everyone else to automate as well.

The result may not be democratisation. It may be an arms race.

D. Human expertise may be displaced from the wrong part of the workflow

AI could help science enormously: literature synthesis, error checking, data analysis, hypothesis generation, code review, replication support, and discovery. But if deployed mainly to generate more papers, it consumes human expertise rather than amplifying it.

The worst outcome is not that AI writes papers. The worst outcome is that human researchers spend more time filtering AI-generated output and less time doing original, difficult, risky, meaningful science.

E. Publishers may face reputational and legal exposure

If journals become known as channels for AI-generated junk, publishers face reputational damage. In high-stakes fields, there may also be legal and regulatory consequences if unreliable research influences clinical practice, product claims, public policy, or professional guidelines.

Publishers will increasingly be judged not only on publication volume, impact factor, or processing speed, but on their ability to prove integrity.

F. The definition of “research contribution” will have to change

If AI can generate endless competent-looking analyses, the sector must stop treating the paper as the primary unit of value. The future unit of value may need to become verified contribution: data, methods, code, replication, negative findings, expert synthesis, clinical utility, reproducibility, and demonstrable novelty.

In other words: less “Can this be written up?” and more “Does this deserve to enter the scientific record?”

5. What the sector should do

There is no single fix. The sector needs a layered response: technical, editorial, economic, institutional, and cultural.

1. Move from detection to provenance

The article quotes the idea that publishers may need to let authors demonstrate authenticity rather than trying to detect fabrication after the fact. That is exactly right.

Detection alone will fail because AI-generated papers will become too polished. The better approach is provenance: authors should increasingly be required to provide evidence of how the research was produced.

This could include:

The burden should shift from “Can the editor catch fraud?” to “Can the submitter prove research integrity?”

2. Treat public-dataset papers as a high-risk category

Public datasets are valuable, but they are now being used as industrial raw material for low-quality publication generation. Journals should introduce stricter rules for papers based on heavily mined public datasets.

For example, such papers should be required to show:

A paper that merely discovers another correlation in a public dataset should not automatically qualify as a research contribution.

3. Build AI-era editorial triage systems

Publishers need stronger triage before peer review. This should not be framed as “AI detection,” because AI detection is unreliable and increasingly beside the point. It should be framed as research-integrity risk scoring.

Signals might include:

The goal is not automatic rejection. The goal is to route suspicious or low-value submissions away from scarce expert reviewers.

4. Protect peer review as scarce expert labour

The sector must stop treating peer review as an infinite free resource. Reviewer fatigue is not an inconvenience; it is a systemic risk.

Possible remedies include:

AI may assist reviewers, but it should not replace expert judgment. AI tools should be used to check references, flag retracted papers, inspect statistical consistency, compare claims with methods, and identify missing reporting requirements. They should not be allowed to produce opaque review decisions.

5. Require stronger AI-use disclosure

AI-use disclosure should become more precise. A vague statement that “AI was used to improve language” is no longer enough.

Authors should disclose whether AI was used for:

This is not about banning AI. It is about understanding which parts of the epistemic chain were machine-assisted and which human authors take responsibility for.

6. Reform incentives beyond publishing

The article’s strongest diagnosis is that the publication-count economy drives the behaviour. Therefore, publishers alone cannot solve the crisis.

Universities, funders, hospitals, regulators, and ranking bodies must reduce reliance on crude publication counts and citation metrics. Assessment should focus more on quality, openness, reproducibility, data contribution, software contribution, replication, societal value, teaching, mentorship, and long-term research programmes.

As long as career advancement depends heavily on producing more papers, AI will be used to produce more papers.

7. Create “knowledge contribution” thresholds

Journals should become more explicit about what counts as a publishable contribution. The question should not be whether a manuscript is coherent. The question should be whether it adds meaningful knowledge.

That means raising thresholds for:

This may reduce publication volume. That should be considered a feature, not a failure.

8. Build shared infrastructure across publishers

No individual publisher can solve paper mills, AI-generated submissions, reviewer fraud, citation manipulation, and provenance gaps alone. The sector needs shared infrastructure.

That could include:

The article suggests the problem is now outgrowing conventional detection models.

9. Separate AI for discovery from AI for paper production

The sector should not become anti-AI. That would be both unrealistic and counterproductive. AI can help science. But the sector must distinguish between AI used to deepen inquiry and AI used to inflate output.

Good uses include:

Bad uses include:

The remedy is not “no AI.” The remedy is governed AI.

6. What this means for scholarly publishers

For scholarly publishers, the article should be read as a warning that the traditional publishing value proposition is shifting. Speed, scale, and volume are no longer enough. In an AI-saturated environment, the value of publishing lies in trust, filtering, provenance, curation, correction, and accountability.

The publisher of the future will not merely host papers. It will need to prove that a paper deserves to be part of the scientific record.

That means publishers should invest in:

This also has commercial implications. If AI systems are trained or grounded on scholarly literature, trusted publishers will need to distinguish their content from polluted or low-integrity corpora. Provenance becomes not just an ethical obligation, but a strategic asset.

7. Conclusion

The article’s central message is uncomfortable but correct: AI is not merely creating fake papers. It is exposing a research system that already had the wrong incentives. The sector built prestige, promotion, funding, and sometimes revenue around publication volume. AI now makes publication volume cheap.

The crisis is therefore not only technological. It is institutional.

The remedy must be equally structural. Publishers need better triage, provenance, AI-use disclosure, shared integrity infrastructure, and stronger editorial thresholds. Universities and funders need to stop rewarding crude output metrics. Researchers need credit for quality, replication, data, software, review, and long-term contribution. AI tools should be used to strengthen verification, not flood the system with plausible noise.

The question at the end of the article — “Do we need more papers?” — is the right one. Science needs more knowledge, more replication, more trustworthy evidence, and more meaningful discovery. It does not need an infinite paper machine.