What does it mean for a writer, such as a novelist, to have a unique “voice”? And does artificial intelligence (AI) help or hurt that voice?
Microsoft researchers set out to answer that question with a small study using 19 fiction writers, 30 readers, and short passages written with the help of OpenAI’s GPT-4. The research takes its title from a comment by one of the writers — “it was 80% me, 20% AI.”
Also: 5 ways ChatGPT can help you write essays
What prompted the study are “concerns that vast transformations of the writer economy are likely underway” as a result of generative AI, writes lead author Angel Hsing-Chi Hwang of the University of Southern California, who collaborated with five scholars from Microsoft Research Montréal. For an author, “authenticity often determines the value of their work, which co-writing with AI might potentially threaten,” the researchers say.
To better understand what “authenticity” means, Hwang and colleagues interviewed the writers from June to October 2023 about their notions of the term, posing questions like, “What are the unique characteristics (tones, phrases, styles, voices, etc.) that make your writing unique?”
They then had each writer use a program called CoAuthor, designed by researchers at Stanford University in 2022. CoAuthor is an interface to a large language model (LLM) that lets a person request and insert suggestions from the LLM as they write by tapping the TAB and ENTER keys on the keyboard.
Also: Don’t trust ChatGPT Search and definitely verify anything it tells you
The 19 participants in Hwang & Co.’s study went through two versions of using CoAuthor to write a 200-word passage. One version was not personalized, using only the native abilities of GPT-4. The other version had some of the individual writer’s work fed into the model at prompt time, an example of what’s called “in-context learning.” The authors did not know which session was personalized and which was not.
Writer concerns
After each of the two exercises, the writers were interviewed about their experience, answering questions like, “Does co-writing with AI affect your authenticity in writing?”
The study found substantial skepticism and even anxiety amongst writers about AI-assisted writing tools.
<!–>
–>
The outline of Microsoft’s study methodology.
Microsoft
As Hwang and team note, authors have repeatedly expressed fears in public discussions and research studies that “using AI might negatively impact their writing outcomes” by “distracting them from their original ways of writing, leading to lower quality of work,” or that “working with AI might diminish their control and joy during the writing process.”
In the present study, the same skeptical attitudes about a negative impact pop up among writers. Hwang notes authors make anecdotal remarks, such as, “Can they get inspiration from AI? Absolutely. But authentic writing is about the story that’s told from the heart of an author.”
Also: The best AI tech of CES 2025: Neural wristbands, smart mirrors and more
Hwang and team define this as the “authentic self” version of authenticity, and they note that the single biggest concern of writers in that respect is “whether a writer can justify having their name and identity behind their work. In other words, a writer claims the authenticity of their writing if they can soundly argue why the piece of work can only be done by them as the writer.”
Despite that, when the writers were told which of their passages they had created with a personalized version of GPT-4, they generally expressed a preference for the one with the personalization. “Overall, in our study, the majority of writers preferred working with personalized writing tools when they were asked to compare the two options,” they write. “This is because writers believed personalization could help preserve their genuine voice, express themselves naturally, and better connect with their own identities.”
<!–>
Writers remarked that they were able to “produce better quality of work under time constraints when working with personalized AI,” and that they felt they were “collaborating” with the personalized AI, versus having to “supervise” the non-personalized AI.
“On the negative end, participants worried personalization might also lead to writers adopting more suggestions from AI, allowing more influences from the tool,” the study observes. “Based on their subjective reflections, many participants believed that they contributed much more to the written content when working with a non-personalized AI, as they more frequently experienced the need to revise content generated by non-personalized AI, leaving more limited room for AI to influence them.”
Experienced writers also expressed concern that novices who had yet to establish their style or voice would simply accept most of the suggestions from a personalized AI system.
Reader reactions
Readers, on the other hand, didn’t really seem to care much.
Out of the 19 writers, Hwang and team selected six representing different genres, including sci-fi and poetry, and had their passages, as well as their original work, read by readers recruited from Reddit. The Reddit participants were “avid readers who regularly read the particular literature genres” of the six writers and were not told which pieces of writing were done with AI.
The readers were asked to rate how much they liked the writing on a scale of one to five, and, for each piece of writing, if they thought it was written “definitely with AI,” “probably,” “probably or definitely written independently by a human writer,” or “No idea.”
–> <!–>
Sample responses from writers about how they define authenticity for themselves.
Microsoft
“Throughout the study, writers expressed concerns about audiences’ reactions to their use of AI assistance for their writing,” the authors note.
However, the survey results indicate readers didn’t find that much difference in the writing samples. “By contrast, readers in our study held a more positive view toward the use of AI writing assistance,” the study notes.
Between the original human-written passages, personalized, and non-personalized AI passages, “the degree of enjoyment […] likeability […] and creativity after reading each type of passage showed no significant difference across the three conditions,” Hwang and team write. And, in a surprising twist: “After being told that some of the passages they read were co-written with AI, readers expressed significantly more positive perceptions toward the writing compared to the median value.”
Readers generally gave the human-written samples more likelihood of being human-written, but the personalized and non-personalized conditions didn’t get any significant difference in scores from the readers.
The authors conclude that “some readers might be able to tell if a piece is co-written with AI with some degree of confidence,” but that “they may not respond differently to content written by the writer alone or content co-written with AI in terms of how much they enjoy reading it.”
The takeaway
Hwang and team ultimately determined that AI-based writing tools need to evolve to offer writers more than they currently do.
“Specifically, our study reveals that authenticity is multifaceted, and the current form of AI assistance (i.e., text suggestion for sentence completion) only contributes to preserving writers’ unique styles and voices in writing,” they say. “Besides text production, writers seek more diverse support (such as practicing externalizing their internal experiences, receiving feedback, and projecting possible audiences’ reactions) to jointly preserve authenticity in their work.”
As a writer, I find movement, such as taking a simple walk, greatly helps the process. But embodiment didn’t seem to come up much for the writers. “Our writer participants generally agreed that most of the AI writing assistance focused on the process of writing, while they hoped to see more options to support relevant activities besides writing per se,” Hwang told me.
Suggestions for improving programs like CoAuthor include personalized recommendations for what an author might read and “offering support during the early stage of idea development (i.e., the ‘fuzzy area’ that writer participants referred to),” which could be “more effective in helping writers preserve their authenticity in writing compared to providing support later on during text production.”
As for the role of tackling the basics of writing, like grammar and syntax, writers were split. “Some believed if AI or any other tools could help them with fixing grammar and other aspects of writing that were less personal, writers could free up some of their capacity and hone in more on crafting their individuality in their work – hence producing more authentic writing,” Hwang says. “On the other hand, some writers saw values in working through these less personal aspects of writing. Sometimes, the greatest ideas occurred when one’s working on the most boring and irrelevant stuff.”
Research blind spots
The researchers acknowledge there are a lot of limitations to their study. I can see one very large issue: For a writer, sitting down to write is very different from an exam-style session where one works with a program such as CoAuthor.
In fact, the authors point out the timed nature of the exercise as an issue for future research.
“Among the three versions of work produced by our writer participants, the pieces composed without any AI assistance were done without any time constraints,” they note. “All writer participants acknowledged that writing under time pressure was a common part of their day-to-day job and that the speed they wrote during the study sessions was at their usual pace.”
