The quality of what arrives is noticeably worse.
In the quiet architecture of scientific knowledge, a new kind of disorder has taken root. Academic journals are now documenting what many have quietly feared: generative AI is reshaping the scholarly record faster than the norms meant to govern it can respond, and the work arriving through this channel is measurably weaker. The disclosure policies journals erected in haste have not held, leaving peer review — already a strained and human institution — to absorb a flood it was never built to filter.
- A leading academic journal has put numbers to a growing suspicion: AI-generated submissions are rising sharply, and the quality of published research is declining alongside them.
- Disclosure policies meant to require transparency about AI use are being ignored, misread, or quietly rationalized away by researchers who may not see themselves as violating anything.
- Peer reviewers — already overextended and unpaid — are not equipped to distinguish machine-assembled prose from genuine scholarly reasoning, letting flawed work slip into the permanent record.
- The telltale signs are accumulating in plain sight: phantom citations, logical gaps, and phrasing no careful human writer would choose, now embedded in journals that once stood for rigor.
- Publishers face a narrowing set of options — detection tools that can be gamed, penalties that require enforcement, or a deeper reckoning with what authorship and integrity mean in an age of generative AI.
Somewhere in the machinery of peer review, something has broken. A leading academic journal recently measured what many researchers have long suspected: artificial intelligence is entering the scholarly record at a pace existing safeguards cannot contain, and the quality of what arrives is noticeably worse.
The findings are straightforward and troubling. Researchers are submitting papers generated or substantially shaped by large language models without disclosing it. The policies journals hastily erected — asking authors to declare AI use — are not working. Some ignore them entirely. Others interpret them loosely. The result is a growing body of published work bearing the fingerprints of machine generation: awkward phrasings, logical gaps, and citations to papers that do not exist.
This is not a marginal problem. When researchers use generative AI without disclosure, they introduce a new kind of noise into the scientific record — work that looks legitimate enough to pass initial screening but lacks the rigor and originality peer review is supposed to guarantee. A researcher might use an AI tool to draft a section and convince themselves it doesn't constitute 'use' in the policy's sense, or assume enforcement is impossible and the requirement merely performative.
The peer review system was never designed for this. Reviewers are already stretched thin, reading papers in their spare time without compensation, and are not equipped to detect machine-generated prose — especially as the machines improve.
What happens next matters. Publishers could invest in detection tools, though these are imperfect. They could enforce penalties, though that requires catching violations. They could demand stricter attestations, though those depend on honesty. Or they could accept that the boundary between human and machine authorship is blurring and rebuild the system around that reality. For now, the evidence is documented. What remains to be seen is whether the institutions that depend on the integrity of the scientific record will treat this as a crisis — or as one more friction point in a system already groaning under its own weight.
Somewhere in the machinery of peer review, something has broken. A leading academic journal recently took the time to measure what many researchers have suspected for months: artificial intelligence is pouring into the scholarly record at a pace that existing safeguards cannot contain, and the quality of what arrives is noticeably worse.
The journal's findings are straightforward and troubling. Researchers are submitting papers generated or substantially shaped by large language models without saying so. The disclosure policies that journals have hastily erected—the ones that ask authors to declare their use of AI tools—are not working. Some authors ignore them entirely. Others interpret them loosely. The result is a growing body of published work that bears the fingerprints of machine generation: awkward phrasings that no human would choose, logical gaps that slip past tired reviewers, citations to papers that do not exist.
This is not a marginal problem. The volume is real. The journal measured it. The degradation in quality is measurable too. When researchers use generative AI without disclosure, they are not simply failing to check a box on a form. They are introducing a new kind of noise into the scientific record—work that looks legitimate enough to pass initial screening but lacks the rigor, originality, and careful reasoning that peer review is supposed to guarantee. A paper written by a human researcher who has thought deeply about a problem reads differently from one assembled by an algorithm trained on patterns in existing text.
What makes this particularly acute is that the policies themselves have failed. Journals issued guidelines. They asked authors to be transparent. And yet the undisclosed use continues. This suggests one of two things: either researchers do not believe the policies will be enforced, or they do not fully understand what they are doing. Perhaps both. A researcher might use an AI tool to help draft a section, then convince themselves that this does not constitute "use" in the way the policy means. Or they might assume that everyone does it, that enforcement is impossible, that the disclosure requirement is performative.
The peer review system was never designed for this. Reviewers are already stretched thin, reading papers in their spare time, often without compensation. They are not equipped to detect whether a passage was written by a human or generated by a machine—especially as the machines improve. A reviewer might flag a paper for poor writing, but by then it is already in the system, already consuming attention and credibility.
What happens next matters. Publishers could invest in technical detection tools, though these are imperfect and can be gamed. They could enforce penalties for undisclosed use, but enforcement requires catching violations, which requires resources. They could demand more rigorous author attestations, but attestations are only as good as the honesty behind them. Or they could accept that the boundary between human and machine authorship is blurring, and rebuild the system around that reality—perhaps by requiring explicit disclosure of which parts of a paper were AI-assisted, the way a researcher might acknowledge a collaborator.
For now, the journal has documented the problem. The evidence is there. What remains to be seen is whether the institutions that depend on the integrity of the scientific record will treat this as a crisis that demands a response, or as one more friction point in a system already groaning under its own weight.
Notable Quotes
Researchers are submitting papers generated or substantially shaped by large language models without saying so.— Journal findings
The Hearth Conversation Another angle on the story
When you say the journal "measured" this, what exactly did they count?
They looked at submissions and published papers and found evidence of AI generation—patterns in language, logical inconsistencies, citations that don't exist. The volume of it was significant enough to be statistically clear.
But couldn't a human write awkwardly? Couldn't a human make citation mistakes?
Of course. But there's a difference between occasional human error and the systematic patterns that emerge when an algorithm generates text at scale. The journal saw enough of a signal to be confident something structural had changed.
Why aren't the disclosure policies working? Are authors just lying?
Some probably are. But I think many authors genuinely don't see what they're doing as "using AI" in the way the policy means. They use it to brainstorm, to draft a paragraph, to help with phrasing. It feels like a tool, like spell-check. By the time they submit, they've convinced themselves it's fine.
What's the actual harm if the science is still sound?
The harm is that you can't tell if the science is sound. A reviewer can't easily spot where the thinking got lazy, where the machine just assembled plausible-sounding text instead of doing the actual intellectual work. And over time, if enough papers are like this, the whole record becomes less trustworthy.
So what would actually fix this?
That's the hard part. You could try to detect it technically, but detection lags behind the tools. You could punish people, but that requires catching them. Or you could accept it's happening and rebuild the system to account for it—maybe require authors to say exactly which parts were AI-assisted, the way they'd acknowledge a collaborator.
Which do you think will happen?
Probably some combination of all three, and probably too slowly. The system is already under stress. Adding another layer of enforcement or detection will be expensive. But ignoring it means the scientific record gets noisier and less reliable. That's a cost too.