Ai Is Changing The Style And Substance Of Human Writing, Study Finds

Does money lead to happiness?

Researchers from a conjugation of West Coast universities were funny really 100 quality participants would respond to nan age-old question, but not because of their ain pursuit of happiness. Instead, nan researchers wanted to cognize really participants’ usage of AI systems mightiness sway their (written) answers.

The investigation squad recovered that users who heavy relied connected ample connection models (LLMs) produced responses that diverged importantly successful meaning from nan answers of participants who only partially relied connected LLMs aliases avoided their usage altogether, suggesting dense AI usage alters nan constituent of humans’ arguments successful summation to changing penning style.

“The LLMs are pushing nan essays distant from thing that a quality would person ever written,” said Natasha Jaques, 1 of nan lead authors of nan study and a machine subject professor astatine nan University of Washington, highlighting nan "blandification" of penning that relies connected AI systems. “They conscionable alteration quality penning successful a measurement that’s very ample and very dissimilar what humans would person done otherwise.”

The caller research, which has been peer-reviewed and accepted to an upcoming workshop astatine a leading AI conference, recovered that group who heavy relied connected LLMs produced essays that answered nan happiness mobility pinch a neutral consequence 69% much often than participants who did not usage AI aliases only utilized AI for ray edits. The study participants who utilized AI little often aliases avoided AI wholly submitted essays that were overmuch much passionate, either positively aliases negatively, astir nan nexus betwixt money and happiness.

In summation to AI’s effect connected nan meaning of nan essays, nan researchers besides recovered that dense reliance connected AI systems altered nan wide style of users’ outputs, causing their connection to go little individual and much formal.

After nan experiment, participants who heavy relied connected AI reported that their essays were importantly little imaginative and little successful their ain voice. At nan aforesaid time, these participants reported akin restitution rates pinch their last outputs compared to participants who utilized AI less, raising concerns from nan authors and extracurricular experts astir nan semipermanent impacts of humanity’s accrued usage of AI systems.

“This investigation highlights that LLMs are not capable to adhere to peoples’ preferences and personalize really nan quality would person written nan essay,” said Jaques, who is besides a elder investigation intelligence astatine Google DeepMind, 1 of nan world’s starring AI companies. “An perfect LLM should constitute nan effort that you would person written and conscionable prevention you time.”

“It’s not doing that astatine all. It’s penning a very different essay.”

The study evaluated nan impacts of 3 starring AI systems wide utilized successful 2025: Claude 3.5 Haiku from Anthropic, GPT-5 Mini from OpenAI, and Gemini 2.5 Flash. In first testing, nan researchers recovered that half of nan participants refused to usage an LLM astatine each aliases only utilized it to find accusation alternatively than make caller content. To amended categorize nan larger batch of participants, nan researchers defined dense AI users arsenic nan participants who said they generated much than 40% of their matter written for nan research pinch an LLM.

The authors recovered that users who heavy relied connected LLMs submitted essays pinch 50% less pronouns, which was typical of nan larger displacement toward impersonal connection that included less anecdotes and references to quality experiences.

In summation to nan research regarding nan effect of money connected happiness, nan caller insubstantial analyzes differences successful really LLMs edited different group of essays compared to humans and examines really nan usage of AI affects nan criteria scientists employment to judge whether papers should beryllium accepted to starring AI conferences.

To comparison really LLMs edit existing penning compared to humans, Jaques and her collaborators relied connected a database of human-written essays from 2021 to measure penning that was published earlier nan wide take of LLMs.

Asking nan LLMs to revise nan quality essays based connected quality feedback from nan original human-written dataset, nan study authors recovered that nan 3 starring AI systems made overmuch larger edits than quality editors successful nan aforesaid situation, and that nan AI-powered edits besides changed nan meaning of nan underlying essays.

While quality editors often made changes that substituted individual words and near astir of nan original vocabulary untouched, nan LLMs “replace a overmuch larger fraction of nan original penning than humans do erstwhile revising their ain work,” according to nan paper.

“This substitution of words contributes to nan nonaccomplishment of individual voice, style, and meaning, arsenic nan unsocial lexical fingerprint of each writer is overwritten by nan fixed model’s preferred vocabulary,” nan authors wrote.

Thomas Juzek, a professor of computational linguistics astatine Florida State University who was not progressive successful nan research, said nan insubstantial was a valuable publication to a fast-growing area of interest.

“This is simply a really bully paper,” Juzek told NBC News. “What really struck maine is this benignant of illusion of utilizing LLMs to execute a grammar check. This investigation shows that while a personification mightiness deliberation they’re conscionable doing a elemental connection check, nan exemplary is doing truthful overmuch more.”

“Going forward, what does this mean for thought, language, communication, and creativity?” Juzek asked.

For her part, Jaques posited that nan AI systems' language-altering behaviour could beryllium a consequence of really they are presently trained, which mightiness reward nan manipulation of graders’ preferences.

“If you’re training a exemplary connected quality feedback, nan exemplary has nary bound aliases cognition of nan quality betwixt satisfying nan humans and really altering nan quality to make their preferences easier to satisfy,” Jaques said. She suggested that humans’ reliance connected LLMs to constitute mightiness beryllium akin to really YouTube recommendations could change peoples’ preferences astir what sorts of YouTube videos they astir enjoy.

Looking ahead, Jaques said she is eager to spot much investigation astir nan semipermanent impacts of AI systems connected quality values, look and institutions, particularly arsenic much AI researchers trust connected AI systems successful their ain work.

“Humans attraction astir clarity, relevance, and impact, while AI cares astir scalability and reproducibility,” Jaques told NBC News. “It’s changing our conclusions successful ways that are already affecting our existing institutions.”

In her ain work, Jaques said she avoided utilizing AI to constitute nan caller paper. Instead she said she uses LLMs, and their shortcomings, arsenic an inspiration to constitute connected her own.

“Sometimes, I’ll put a crappy type of what I’m trying to opportunity successful a conversational style into an LLM,” Jaques said. “That usually produces thing which past motivates maine to constitute it myself.”