When a researcher uses a large language model to help write a paper, the AI often leaves telltale traces. Words like “delve,” “underscore,” “intricate,” and “pivotal” cluster at rates far higher than in human-written text. Sentence structures follow predictable patterns. Openers like “In recent years…” and transitional phrases like “not just X, but Y” flag the text as AI-generated to trained readers, and to detection software.
Jie Ding, an associate professor at the University of Minnesota’s School of Statistics, has released a tool designed to remove those traces. Called the Academic Humanizer, it is a set of structured prompt instructions that can be fed into an AI coding agent, Claude Code, Codex, or similar, to audit an academic manuscript for AI tells and rewrite it to sound human.
“The job is to strip the AI tells without casualizing, and to enforce the discipline a general humanizer misses: every claim earns its number, figure, or citation, and no verb is stronger than its evidence,” reads the tool’s README on GitHub.
Six layers of cleanup
The Academic Humanizer operates across six layers. First, a general AI-tell catalog removes overused words (delve, intricate, tapestry, pivotal, foster, leverage, realm, seamless, and others). Second, it targets academic-specific tells: over-claiming verbs, significance hype, empty intensifiers, novelty padding, and formulaic openers. Third, it preserves scholarly conventions. Fourth, it matches every claim to its evidence, no unsupported assertions. Fifth, it calibrates voice and venue. Sixth, it offers a funding-proposal mode tailored to NSF and NIH review structures.
A before-and-after example from the README illustrates the effect. Before: “In recent years, continual learning has attracted increasing attention and achieved remarkable success. However, existing methods still face crucial challenges. In this proposal, we propose a novel framework that leverages cutting-edge techniques to delve into these intricate problems, paving the way for a transformative paradigm that will revolutionize the field.” After: “Continual learning matters, but today’s methods stay empirical and their principles are unclear. That limits reliability and progress. This proposal builds a principled framework on three fronts: adaptation, soft supervision, and cross-domain knowledge.”
The ethics debate
Reactions have been sharply divided. Francisco Maria Calisto, a health-informatics researcher at the University of Lisbon, told Nature he uses the tool heavily. “It’s the best I have ever used,” he said, for emails and code documentation as well as manuscripts.
Miguel Angel Blazquez Rodriguez, a plant biologist at the Polytechnic University of Valencia, was dismissive: “I don’t like it. It’s deceiving.”
Cassidy Sugimoto, an information scientist at Carnegie Mellon University, expressed concern: “I fear that the use case is harmful for science. I’m worried.”
Ding himself draws a distinction between the tool and its misuse. “I’d separate the tool from the behavior,” he told Nature. “The ethical issue is the non-disclosure and the intent behind it, not the existence of an editing aid.”
After media queries, Ding updated the tool’s description from “removes the usual AI tells” to “sharpens clarity and voice,” and added an ethics note clarifying that the tool does not remove the obligation to disclose AI assistance.
A detection arms race
The release of the Academic Humanizer comes against a backdrop of rising AI use in academic publishing. A February 2026 study by He and Bu in PNAS (DOI: 10.1073/pnas.2526734123) analyzed 5,114 journals and 5.2 million papers, finding that despite 70 percent of journals adopting AI disclosure policies, AI writing use has surged with no statistical difference between journals with and without such policies. Of 75,000 papers published since 2023, only approximately 76, roughly 0.1 percent, explicitly disclosed AI use. The authors concluded that current policies have “largely failed” to promote transparency.
Detection companies are responding. Pangram Labs, a Brooklyn-based AI detection startup founded in 2024 by former Google and Tesla engineers, tested the Academic Humanizer. CEO Max Spero said initial tests “caught most humanized text, but not all”, and Pangram is designing upgraded versions to detect the tool’s output. In a 2025 preprint (arXiv:2501.03437), Pangram demonstrated a model called DAMAGE that was robust to 19 AI humanizer and paraphrasing tools, claiming more than 90 percent accuracy on all tested humanizers.
The question is whether the arms race between humanizers and detectors benefits scientific integrity or simply raises the cost of dishonesty. For now, the Academic Humanizer remains available on GitHub, and the debate over whether it is a legitimate editing tool or an instrument of deception is far from settled.
Sources
1. Nature News, “‘Humanizer’ tool erases signs of AI-written text” (7 July 2026). DOI: 10.1038/d41586-026-02105-3
2. Academic Humanizer GitHub repository: https://github.com/AIScientists-Dev/academic-humanizer
3. He, Y. & Bu, Y., “Academic journals’ AI policies fail to curb the surge in AI-assisted academic writing”, Proc. Natl Acad. Sci. USA 123, e2526734123 (2026). DOI: 10.1073/pnas.2526734123
4. Masrour, E., Emi, B. & Spero, M., “DAMAGE: Detecting Adversarially Modified AI Generated Text”, arXiv:2501.03437 (2025)