Don’t Use AI To Edit Writing

While teachers largely agree that Large Language Model AI such as ChatGPT or Claude should not be used by students to write entire assignments, there is another school of thought that AI can be used by students to "help" with writing assignments, including the use of AI to edit or "clean up" a student rough draft. A new research paper suggests that's a bad idea.

" How LLMs Distort Our Written Language " shows that LLM-based revision results in "large changes in the resulting content and meaning." It's not simply that LLMs convert language to a sort of bland linguistic average; in many cases, the researchers found that the LLM changes the writer's intended meaning.

Heavy users of LLM reported that their essays "do not reflect their own voice," and researchers further show that LLMs distort language into one particular LLM "style," making it semantically unlike human-generated essays.

Asked to write an essay on the question "Does money lead to happiness," the essays either "assisted" or revised by the bots were more likely to take a neutral stance, a "both sides" approach to the question. Even when prompted to stick to fixing grammar, the bots still changed the original stance of the author, moving it to the middle position in a way that humans revising their own work did not. Similar effects were demonstrated with an essay supporting or opposing self-driving cars.

Many AI users have noted, as writer Alba Tay pointed out, " AI will always say you're right. " That positive slant was observed in this research "with LLMs expressing slightly more words relating to positivity and trust labels in the essay," putting a positive slant on any point of view included in the essay.

The researchers conclude that "LLMs significantly alter the meaning of human writing," and this includes instances in which the bot is told strictly check for structural issues like grammar, usage, and spelling.

The researchers also assert that the bots favor "complex, formal constructions over the more direct, personal style typical of human writing," tending to create longer sentences using more coordinating conjunctions, more nouns, and fewer pronouns. They tend to remove first person voice from the writing.

They express wider concerns, noting that "even though people who rely heavily on AI recognize that it diminishes their voice and creativity, they are nevertheless equally satisfied with the results." What happens in society when writers are perfectly happy to sign off on writing that imperfectly represents their actual thoughts?

That points to all sorts of troubling implications for culture and institutions, but it is a particularly huge red flag for LLM use by students. At the very moment when they should be fostering voice and creativity and learning ways to say exactly what they mean, using LLMs, even as editing tools, seems likely to lead students in exactly the wrong direction. Teachers considering allowing students use bots "just for editing" should reconsider.

It's not simply a question of cheating or cutting corners, but of using a tool that undercuts the very purpose of writing itself. If writing is not built to bridge the communication gap between human beings, then what is it for?

The paper was co-authored by Marwa Abdulhai , Isadora White , Yanming Wan, Ibrahim Qureshi, Joel Z. Leibo , Max Kleiman-Weiner , and Natasha Jaques , academics representing UC Berkeley, UC San Diego, University of Washington, Zaytuna College, and Google DeepMind.

The main text of the paper is sixteen pages, and while some of the data crunching is very academic, the heart of the paper is easily accessible for the civilian reader. Teachers who are wrestling of the question of just how far LLMs can be trusted to aid students should take a moment to read this research.

This article was originally published on Forbes.com