Machines are increasingly being used to edit written output.Credit: Getty
A new academic ‘humanizer’ tool aims to personalize the tone of research papers written with an artificial-intelligence program, in part by erasing apparent signs of AI usage. Some researchers praise the tool — but others are voicing concerns.
The tool, released on 20 June, is tailored for “papers and grant proposals”, according to its developer, Jie Ding, a machine-learning researcher at the University of Minnesota in Minneapolis. It can be used to “match the author’s own voice”, according to the GitHub site for the tool, which lists under “core principles” the need to “strip the AI tells without casualizing”.
The new humanizer is “not sophisticated,” says Max Spero, chief executive of a firm in New York City that produces an AI-detection platform called Pangram. In his initial tests of ‘humanized’ text, Pangram caught most of the AI-generated language — although not all of it. Spero says that upgraded versions of Pangram are being designed specifically to detect humanizer use.
Scientists are increasingly turning to AI systems to help them write papers, grant applications and even peer reviews. AI assistance can be a boon, particularly to people writing in a language that is not their first, and many publishers allow some degree of AI use in preparing papers1 if such use is declared.
AI linked to explosion of low-quality biomedical research papers
Some rave reviews of the humanizer are from scientists. “I am using the tool a lot,” mostly to assist with writing e-mails and documentation for code, says Francisco Maria Calisto, a health-informatics researcher at the University of Lisbon. “It’s the best I have ever used.”
Other researchers say humanizers are not good for science. “I don’t like it,” says Miguel Angel Blazquez Rodriguez, a plant biologist at the Polytechnic University of Valencia in Spain. “It’s deceiving.”
Rodriguez and others fret that humanizers will tempt more scientists to use AI without disclosure. “I fear that the use case is harmful for science,” says Cassidy Sugimoto, an information scientist at Carnegie Mellon University in Pittsburgh, Pennsylvania. “I’m worried.”
Asked about these concerns, Ding responded, “I’d separate the tool from the behaviour,” adding that failing to disclose AI assistance when required is misconduct, regardless of how a piece of text was produced. “The ethical issue is the non-disclosure and the intent behind it, not the existence of an editing aid.”
After being asked about the potential for deceit, Ding made updates to the GitHub site, such as changing a description of the tool’s function from “removes the usual AI tells” to “sharpens clarity and voice”. He also added an ethics and disclosure note, which he says clarifies “that the tool is an editing aid and does not remove the author’s obligation to disclose AI assistance.”
AI scribes
Humanizers are finding increasing favour among researchers who use AI, Calisto says, because different AIs all “sound the same”. The tone of AI-generated text is sometimes inappropriate for academic writing, says Ding. For example, the tools can exaggerate the strength of a scientific claim, he says. And most AI output has a similar style, which readers tire of, he adds.
‘Obviously ChatGPT’ — how reviewers accused me of scientific fraud
Ding adapted a previously developed humanizer for academic content. His version tells an AI to remove its ‘tics’ from a text and to add more evidence for unsupported scientific claims. Anyone can copy Ding’s guidelines into their AI and tell it to apply the rules, which include avoiding phrases of the form “not just X, but Y” and removing em dashes. Both of those constructions have become associated with AI, according to the tool’s website.
