Is AI making everyone’s writing the same?

A new USC study suggests that it may lead to more uniformity — and loss of individual voice — in written language. (Photo/Pexels)

Remembering the human: Encouraging unique voices in the age of AI

As a USC study highlights the influence of AI tools on linguistic variety in writing, faculty share how they teach students to engage with artificial intelligence in ways that stimulate — instead of stifle — their creativity.

As students begin a new school year, USC experts are working to understand artificial intelligence’s effects on the classroom, especially on creative work such as writing.

Faculty members are not only studying how AI tools such as ChatGPT have affected written language. They are also learning how educators at all levels can help students navigate the world of AI and leverage tools in ways that improve creativity instead of impairing it, said Stephen J. Aguilar, associate professor at the USC Rossier School of Education.

“Teachers have to have a policy even if schools don’t,” said Aguilar, who leads USC’s Center for Generative AI and Society and researches how educational technologies influence teaching, learning and motivation. “How do you want students to use it? Understand its usage; for instance, know how your assignments work when put into popular chatbots and become familiar with AI output. Develop assignments that have students think critically about AI, engage with the tool and interrogate the output together. When you have that engagement, students think about AI in a much more sophisticated way.”

Study: Increased conformity in papers, posts and more

As more people let AI tools do the writing for them, a new USC study suggests that it may lead to more uniformity — and loss of individual voice — in written language.

The research, published in Nature Human Behaviour, argues that large language models (LLMs) — the type of artificial intelligence system powering tools such as ChatGPT — have narrowed the range of complexity and personalization found in writing, leading to increased homogenization of writing styles.

“Language isn’t just a communication tool; it actively complements and shapes our inner reasoning,” said senior author Morteza Dehghani, professor of psychology and computer science and director of USC’s Cognition, Alignment and Language in Machines (CALM) Lab. “When AI systems standardize human expression, they risk flattening the diverse cognitive landscapes that drive collective intelligence. Shrinking our linguistic variation doesn’t merely change how we write; it threatens to narrow the very ways we justify and reason through ideas.”

Researchers in the CALM Lab analyzed more than 880,000 pieces of publicly available text, ranging from academic papers and articles in local news outlets to posts on the popular social discussion platform Reddit.

The team quickly noticed that the range of text complexity narrowed over time, and this trend was first noticeable in November 2022 — coinciding with the launch of ChatGPT.

“The finding isn’t that writing got simpler; what was reduced was the diversity,” said first author Zhivar Sourati, a USC doctoral candidate in computer science. “Some people naturally write in a short, direct way, while others use rarer words and more elaborate sentences. After LLM rewriting, those different styles moved closer together.”

While the meaning behind each piece of text was largely preserved after being rewritten by the LLM, the narrowing of language diversity extended beyond complexity.

Writers’ use of pronouns such as “I” and “we,” conversational phrases such as “you know,” agreement words and emotional language also became more uniform. Emotional expression was not erased — in fact, rewritten texts contained slightly more emotional language on average — but writers began expressing emotion in similar ways, Sourati said.

These changes made it harder to detect the subtle cues about personal attributes typically embedded in writing, Sourati said. Computer models designed to use clues in text to predict personal traits became about 6% less accurate, on average, at predicting authors’ age range, political affiliation, personality, empathy and moral values after their writing was rewritten by an LLM.

Some known associations — such as those between pronoun use and extraversion, friend-related words and loyalty, or complex words and openness — grew weaker, while others remained. Rewritten texts consistently sounded more like they had been written by someone older, male, politically liberal, more morally concerned and less empathic, Sourati said.

The effect varied by setting, with creative writing on Reddit losing the most variation, followed by academic abstracts; news writing was more resistant, possibly because newsrooms already enforce consistent styles, he added.

AI literacy that strengthens student voices

USC Writing Program faculty members say these findings highlight why they’re using new educational approaches to help students navigate AI tools in a more thoughtful way that promotes critical thought and creativity.

“Studies like this underscore what we have been telling our students in Writing 150 and 340 since ChatGPT burst onto the scene: High-level critical thinking and writing skills will be even more important in the future than ever before,” said Daniel Pecchenino, associate director of the USC Writing Program and professor (teaching) of writing at the USC Dornsife College of Letters, Arts and Sciences. “I believe that the challenges presented by generative AI offer us tremendous opportunities to innovate pedagogically, both by designing and critically engaging with AI tools, and through purposeful disengagement from them. Writing faculty, who have always been leaders in teaching, are uniquely positioned to help lead campuses toward a future where AI doesn’t make education less challenging and valuable, but actually more rigorous and meaningful.”

Along with incorporating AI literacy information into core writing classes, the new USC Dornsife Intensives include a three-course “Writing and Responsible AI” credential that is open to USC undergraduates in any major. The end goal is for students to learn how to avoid the potential pitfalls of AI and leverage these tools to bolster their own skills and creativity in a changing world.

“In this first wave of generative AI, our students need ‘resistance training’: resistance both in the sense of pushing back against all the AI hype and in the sense of building intellectual strength through critical and creative exercises with AI,” said Mark Marino, professor (teaching) of writing at USC Dornsife. “We are giving them that training through lessons in AI history, ethics and analytical thinking. Ironically, as these technological systems generate polished but hollow prose, students are rediscovering the value of traditional skills like close reading, the authority that comes from their lived experience and the richness of their own unique voices.”