Writing an email with AI may be easy. Writing a wedding vow with AI—and making it still sound like you—is another matter.
A new University of Maryland study led by fifth-year computer science doctoral student Connor Baumler found that editing AI-generated text can make it feel more personal, but the technology’s stylistic fingerprints often remain. The research highlights a challenge for people turning to AI for writing where personal voice matters most.
The study was presented last month during a poster session at the 2026 Annual Meeting of the Association for Computational Linguistics (ACL) in San Diego.
To explore how well AI-assisted writing can reflect a person’s voice, the research team asked 81 volunteers to complete writing tasks in which personal style was especially important. Researchers first generated drafts using participants’ key ideas, then asked the volunteers to edit those drafts to better match their natural writing styles.
The participants generally felt the revised drafts accurately reflected who they were and said they would consider using the workflow again. But computational analysis revealed a disconnect: Even after substantial editing, the final pieces remained stylistically closer to AI-generated writing than to the participants’ independently written work.
Baumler cautions against treating those computational measures as the final word on whether a piece of writing sounds authentic.
“Because we’re so concerned about this personal writing and personal sense of style, it’s very much possible that the automated metrics are failing to capture what the person really thinks is important about their style,” Baumler explained.
That disconnect also raises questions about how AI systems should be evaluated—and how much authority researchers should give automated measures of personal style.
“I think it’s important as AI researchers to not overstep and prescribe to people that the automated metric says ‘This doesn’t sound like you’ or imply that the model knows what you sound like better than you do,” Baumler said.
The distinction is especially important as AI moves from routine workplace tasks into more personal forms of communication. Baumler, whose broader research focuses on human-AI interaction and fairness, said people may have very different expectations for AI depending on what they are writing.
“If I’m using an AI system to write an email to my adviser or write a paper, it is a little less personally important that these stylistic things sort of match up with my own personal sensibilities,” Baumler said. “But this personal writing setting is an interesting area where people feel differently about what the appropriate use cases are.”
The study draws on the expertise of Baumler’s co-advisers, Computer Science Professors Hal Daumé III and Marine Carpuat. Both hold appointments in the University of Maryland Institute for Advanced Computer Studies (UMIACS) and are members of the Computational Linguistics and Information Processing (CLIP) Lab.
“People express so much about themselves through what they write—not just what they say, but how they say it,” said Daumé, who is also the director of the Institute for Trustworthy AI in Law & Society (TRAILS). “Our voice is such a core part of who we are, and making sure that voice comes through—especially in emotionally charged interactions like apologies or wedding vows—is part of what maintains our humanity.”
The study grew out of an opportunity to apply existing AI analysis tools to human-AI co-writing. The team used metrics developed through the IARPA HIATUS project, which partially funded the work. Carpuat served as the UMD principal investigator for the initiative.
Baumler credits his co-advisers with helping shape the project’s direction. Carpuat’s background in cross-cultural and human-centered communication helped ground the study in how people actually use language technologies, while Daumé’s focus on the societal impacts of AI aligned with Baumler’s interest in human-AI interaction and its effects.
As generative AI becomes more deeply embedded in everyday life, the study raises questions about authorship, authenticity and how writing tools should support human voices without smoothing them into something less personal. Rather than focusing only on technical benchmarks, Baumler hopes researchers will pay closer attention to what users actually want from AI systems.
“I would hope that AI researchers have a little more consideration for what it is that the user really wants to get out of the system,” Baumler said. “What are the appropriate places to apply AI, versus just going for a very blanket, ‘We have this tool that can do everything and therefore it should do everything.’”
—Story by Diya Sharma, UMIACS communications group