A tool strips AI tells from research papers: flagged words, em dashes, stock phrasing. Detectors read the structure underneath, and editing does not change it.
There is now a tool that removes the signs of AI from scientific text. It came out on June 20, built for research papers and grant proposals. On July 7, Nature wrote about it: some scientists like it, others call it cheating. The tool deletes typical AI words and em dashes. It does not change the statistical structure of AI-generated text. Detectors read that structure, not the word list. If you review manuscripts, this difference is now part of your job.
هل تريد معرفة نتيجة نصك؟ تحقّق منه هنا مباشرة.
The tool comes from Jie Ding, a machine-learning researcher at the University of Minnesota. It is a set of instructions for a language model, published on GitHub. The main rule is quoted right in the file: "strip the AI tells without casualizing." The tool removes words like "delve" and "tapestry." It removes the "not just X, but Y" pattern. Even em dashes get deleted. And grant proposals have a mode of their own, NSF and NIH included.
Some of this is ordinary editing, the good kind. A claim with nothing behind it, no number and no citation, gets flagged back to the author, and the paper is better for it.
Scientists disagree about the rest. Francisco Calisto from the University of Lisbon uses the tool often and calls it the best he has tried. Miguel Angel Blazquez Rodriguez from the Polytechnic University of Valencia said one thing: "It's deceiving." Cassidy Sugimoto from Carnegie Mellon said: "I fear that the use case is harmful for science." Ding answers that the tool is not the problem. In his view, the violation is hiding AI help when the rules require disclosure. But after Nature asked questions, the description on GitHub changed. "Removes the usual AI tells" became "sharpens clarity and voice." An ethics note appeared. The tool itself works the same as before.
A manuscript in your review queue may have gone through this tool already. The author will not tell you.
About one third of new arXiv preprints read as machine-written. This comes from a July study by an independent team. They scored the full text of preprints from 2021 to 2026 with a calibrated AI detector. The method is careful. The team set the detection threshold so that pre-ChatGPT papers, which are human by definition, get flagged only 0.4% of the time. The gap between fields is large: close to two thirds in text-heavy fields, near zero in mathematics. The authors also say their numbers are a lower bound, because some machine text passed the detector unnoticed.
Share of new arXiv papers flagged as AI-written, by field
Which leaves anyone reading submissions with one workable question per paper: how much machine text is inside, and did the author admit it.
At Organization Science they got tired of guessing and simply counted. The journal's AI Task Force combed through five years of manuscripts and reviews. Since ChatGPT arrived, submissions are up by nearly half, and when the team traced where that growth came from, almost all of it turned out to be papers thick with AI writing. COVID, for comparison, barely dented the same curve.
Nobody showed the editors any AI scores. They worked the way they always work, reading a paper and deciding whether it deserves reviewers' time, and the papers heavy with machine text still lost. Desk rejections in that group jumped. Very few made it to the revise-and-resubmit stage, while mostly human papers got there about three times as often.
What Organization Science found in five years of its own submissions
One more finding hurts a popular theory, that AI writing gives non-native English speakers a fairer chance. In this data set the opposite happened, the heaviest users came out of review worse off. The report closes with a suitably tired line, "the humans are getting tired."
A humanizer changes none of this. The paper still reads weak, and all the author wins is that nobody catches the word "delve."
arXiv used to trust an institutional email address. Not anymore: a first-time submitter now also needs published co-authorship in the field, or a personal endorsement from an established author. The platform explained itself without much diplomacy, there were simply too many non-scientific submissions to handle. Mathematics had switched to the stricter rule months earlier, and rejection rates on the platform have more than doubled since the flood began.
Because the incentives have not moved. Steven Zhou from Claremont McKenna built a simulation of scholars moving toward tenure, and it confirms what most people in academia quietly know already: when a career depends on how many papers you publish, AI-assisted writing is the smart bet. So if submissions land on your desk, waiting for authors to police themselves is not a plan. Write the procedure.
Consider what an AI humanizer actually is: a language model, editing the output of another language model, with instructions to sound less like a language model. Everything it can reach sits on the surface of the text.
What an AI humanizer changes and what stays in the text
In the early tests Nature describes, most of the cleaned text still got flagged. A sliver did get through.
Now the limits. No independent team has properly tested humanizers against AI detectors so far, meaning every number in this fight comes from a vendor, ours included. Distrust us all equally, that is the healthy default. At It's AI we keep our accuracy numbers on a public page and build the test set around the nastiest case we know, machine drafts edited by a person, because that is the exact thing a humanizer hands you. A detector, by the way, was never a cheater trap. It puts one number on the table, the share of machine writing, and the editor compares it with what the author declared.
If you edit a journal: put the AI rules in writing, who may use what and how they declare it, and give intake a text check the way it long ago got a plagiarism check. Reviewers do this filtering by eye today. It works, and it slowly burns them out.
Authors have it easier. Declare what you used, and run your own text through an AI detector before submitting, the journal will run its own check anyway.
Humanizers will keep improving, detectors will too. A text you are willing to put your name on has nothing to fear from either.
To humanize AI text means to rewrite machine-generated output so it reads as if a person wrote it: varied sentence rhythm, fewer stock phrases, no overused connectors. Academic humanizers go further and strip specific giveaways, from flagged vocabulary to em dashes, while keeping the formal register a research paper requires. The edit changes surface style only. The statistical fingerprint of the language model, such as predictable word choices and even pacing, stays in the text. That is why humanized output is best treated as a hybrid of machine and human writing, not as human prose.
AI writing in a research paper is detected through a mix of statistical scoring and editorial judgment: an AI detector reads the probability patterns of the text, while editors weigh the argument itself. Full-text scanning works better than checking an abstract alone, because a paper can score low on its abstract and high on its body. Reliable checks also report a false-positive rate measured on pre-ChatGPT writing, since formal academic prose by non-native English speakers is the classic false-alarm case. A single score should start a conversation about disclosure, not serve as a verdict on its own.
Often yes, though with reduced confidence. In the first tests Nature reported in July 2026, a detector caught most of the text processed by an academic AI humanizer, but not all of it. Humanizers rewrite the wording layer, while detectors also read deeper signals: how predictable the next word is, how evenly the sentences run. No independent benchmark of humanizers against AI detectors has been published yet, so treat both "undetectable text" claims and "catches everything" claims with the same caution. Testing on hybrid, human-edited AI text is the closest available proxy.
Rankings of AI humanizers answer the wrong question for academic work. What matters is whether the venue you submit to allows undisclosed AI writing, and most journals and funders now require disclosure regardless of how polished the output reads. A humanizer can clean the style of an AI draft, and used openly it is an editing aid. Used to hide AI writing that a policy says must be declared, the same tool turns a style edit into a misconduct case. The venue's disclosure rules, not a tool ranking, decide what is safe.
It depends on disclosure, not on the tool. Editing an AI draft for clarity and declaring the AI assistance is ordinary revision; even the developer of the academic humanizer covered by Nature says the misconduct is hiding AI help where rules require reporting it. Running a manuscript through a humanizer specifically so that an AI detector and an editor see "human-written text" is bypassing detection in the way journal policies mean it. The distinction editors apply is simple: does the declared level of AI use match what is actually in the text.
