A new academic “humanizer” is starting to worry journals, universities and funders. Nature reported in July that an open source humanizer tool, available as a Claude skill, can direct text systems to remove writing patterns linked to AI in research papers and grant proposals. The tool was built for a world where journals, schools and funding bodies treat AI-written prose as a thing to catch, score and punish. Some researchers who struggle with writing well in English say that the tool is meant to make their research more understandable while others say that this is just another form of academic dishonesty.

What The Humanizer Actually Does

The tool at the center of the debate is called Academic Humanizer , an open source project built to edit AI assisted drafts of research papers and grant proposals. The tool, like other AI humanizers , can remove writing patterns linked to AI use in academic work, which is why some researchers see it as another step in the detection arms race. The project’s own GitHub page says it trims generic AI phrasing, long sentences, stock openings, em dashes and phrases such as “paves the way,” “to the best of our knowledge,” “delve” and “underscore.” It can also be pointed at a researcher’s prior work so the new draft sounds closer to that author’s established voice.

The developers frame Academic Humanizer as a clarity and voice tool, not a cheating tool. Its README says it does not create results, invent data or change citations, and that users still have to disclose AI help under journal or funder rules. The Academic Humanizer is built as a Claude skill and can be aimed at a user’s earlier papers.

AI Writing Is Already Part of the Scientific Publishing Process

A Science Advances study examined more than 15 million PubMed abstracts from 2010 to 2024. The researchers found significant vocabulary shifts after large language models arrived and estimated that at least 13.5% of 2024 biomedical abstracts had been processed with LLMs. In some slices of the data, the lower bound reached 40%.

For many researchers, AI is a language tool. It can clean grammar, translate clumsy sentences and help non native English speakers compete in a system where English fluency often gets mistaken for intellectual quality. Used that way, it looks less like cheating and more like a Grammarly style word processor that aims to sharpen a message instead of create one from scratch.

The trouble starts when the same tools fill the page with thin or recycled claims, fake references and paper mill generic content that makes different papers sound the same. Nature reported in April that tens of thousands of 2025 publications might contain invalid AI generated references. A separate 2026 preprint that audited 111 million references in 2.5 million papers estimated 146,932 hallucinated citations in 2025 alone. Editors aren’t bothered when writing style is improved, but they are very concerned when claims don’t hold up to scrutiny.

Companies in the content and publishing industries have all experimented with AI detection tools, with mixed to poor results. AI detection works by spotting machine patterns and tells that indicate that a computer is primarily responsible for text content. However, this approach is brittle and easy to circumvent. Tools on the market developed quickly to outwit AI detection systems, leading to general disuse of detection technology.

Compounding problems of poor detection quality are false positives when AI detectors accuse humans of AI generated content that is indeed their own writing. That is especially dangerous in academic settings, where students, junior researchers and non native English writers may already face unequal scrutiny. False accusations of AI generation can damage careers.

Humanizing Versus Generating

In a high paced publishing and content industry, there is a role for AI in the content generation, improvement, editing and moderation process. However, the balance is tricky to get right. Scientific publication editors need experts to read and check articles, but expert readers are busy. Reviews are often unpaid, rushed and squeezed between teaching, clinic hours, lab management and grant deadlines. In that setting, use of AI in both the content creation and editing processes offers a tempting shortcut.

Likewise, researchers face pressure to publish more, apply for more grants and move faster through a system that rewards volume. AI can help polish language, organize arguments and make dense science easier to read. It can also turn weak work into prose that sounds more convincing than the evidence behind it. This is where the risk sits and where publishers and readers see the problem with AI involvement, in any capacity.

AI-generated content can easily produce unchecked citations, thin data and claims that editors can’t properly vet nor authors fully defend. For editors, a machine written review can sound thorough, but it can still miss the experiment that was never run, the control that makes no sense or the citation that says something narrower than the manuscript claims.

The signs are already here. One study estimated that 6.5% to 16.9% of text in recent AI conference peer reviews was substantially modified by an LLM, with rates varying by venue. Another study of ICLR 2024 estimated that at least 15.8% of reviews were written with AI assistance and found that AI assisted reviews could affect scores and acceptance odds near the decision threshold. an open source humanizer tool, available as a Claude skill, can direct text systems to remove writing patterns linked to AI in research papers and grant proposals.

Even more worrisome, the Washington Post reported in 2025 that researchers had found hidden instructions meant to influence AI assisted review systems. The Guardian covered similar concealed prompts in papers on arXiv.

Publishers And Funders Are Moving, But Slowly

As a result of the increasing spread of AI in publications and the challenges for authors, reviewers and editors, publications are coming up with new rules while also looking to support humanizers that can improve the communication of key insights.

Elsevier says AI tools cannot be listed as authors, since authorship carries duties that only humans can meet. Its journal policy places responsibility for accuracy and integrity on human authors. Springer Nature tells peer reviewers not to upload manuscripts into generative AI tools and says it does not attribute authorship to AI.

Funding agencies have taken a tougher stance on review. NIH prohibits scientific peer reviewers from using large language models or related generative AI tools to analyze and write critiques for grant applications and R&D contract proposals. The National Science Foundation bars reviewers from uploading proposal content or review records to non approved generative AI tools.

Nature reported in May that funding agencies were already battling a wave of AI assisted applications. In addition, big shifts in US and European research are causing changes to how research is prepared and funded. Resource-strapped agencies need to manage volume without building rules that punish less connected researchers or entrench old hierarchies.

The open source humanizer tool matters less as a specific application than as an indication as to where the markets are heading. Clearly there’s demand for tools to help with the process and communication aspects of research that guide funding and support. And on the review side, AI is playing a greater role in the research workflow, shaping drafts, reviews, proposals and editorial triage. Where things need to head is a direct and honest dialogue of where the balance will sit between a complete, and unsuccessful, ban of any AI in the research process, and total openness to AI content tools.

This means scientific publishing has to stop confusing style with trust. The winners will be the institutions that treat AI use as part of the research supply chain and build checks around evidence, provenance and responsibility before the next tool makes detection look even more impossible.