Study Finds AI Is Changing Both the Style and Substance of Human Writing

Summary

New research from a coalition of West Coast universities has concluded that heavy use of large language models (LLMs) does not merely alter the style of human writing — it also changes its content and meaning. The study has been peer-reviewed and accepted for presentation at an upcoming workshop at a leading AI conference.

The research started from a simple question: "Does money lead to happiness?" One hundred participants responded to this question in writing, with varying levels of AI assistance. The results showed that those who relied most heavily on LLMs produced responses that diverged significantly from those who did not use AI or only made light edits with its help.

Natasha Jaques, one of the lead authors of the study and a computer science professor at the University of Washington — also a senior research scientist at Google DeepMind — summarised the phenomenon with a single word: "blandification." According to Jaques, "the LLMs are pushing the essays away from anything that a human would have ever written."

In practice

The study's concrete findings are striking. Participants who relied most heavily on LLMs produced essays that responded neutrally to the money-and-happiness question 69% more often than participants who did not use AI or only used it for minor edits. By contrast, participants who used AI less or avoided it entirely submitted essays that were far more passionate — either positively or negatively — on the topic.

In terms of style, texts produced with greater AI reliance contained 50% fewer personal pronouns, reflecting a broader shift toward impersonal language that included fewer anecdotes and references to concrete human experiences.

The study evaluated the impact of three AI systems widely used in 2025: Anthropic's Claude 3.5 Haiku, OpenAI's GPT-5 Mini, and Google's Gemini 2.5 Flash. In initial testing, the researchers found that half of the participants refused to use an LLM at all or only used it to find information rather than generate new content. For analysis purposes, they defined "heavy users" as participants who reported generating more than 40% of their text with LLM assistance.

The research also examined how LLMs edit existing writing compared to human editors. Using a database of human-written essays from 2021 — prior to the widespread adoption of LLMs — the researchers asked the three AI systems to revise those texts based on human feedback from the original dataset. They found that LLMs made much larger edits than human editors in the same situation, replacing a much larger fraction of the original vocabulary. While human editors tended to substitute individual words while leaving most of the original text intact, the AI models overwrote each author's unique lexical fingerprint with the model's own preferred vocabulary.

After the experiment, participants who relied most on AI acknowledged that their essays were significantly less creative and less representative of their own voice. Yet paradoxically, they reported similar satisfaction levels with their final outputs compared to participants who had used AI less — a finding that concerned the authors and outside experts, given the potential long-term impact of this gap between perception and reality.

Context

Jaques offered a hypothesis about the origin of this behaviour: it may be rooted in how models are currently trained, in ways that reward the manipulation of evaluators' preferences. "If you're training a model on human feedback, the model has no boundary or perception of the difference between satisfying the humans and actually altering the human to make their preferences easier to satisfy," she said. Her comparison is to YouTube recommendations: just as the algorithm can shift users' preferences about what videos they enjoy, LLMs may be shifting users' preferences about what constitutes good writing.

Thomas Juzek, a professor of computational linguistics at Florida State University who was not involved in the research, praised the study and highlighted one point in particular: "What really struck me is this kind of illusion of using LLMs to perform a grammar check. This research shows that while a user might think they're just doing a simple language check, the model is doing so much more."

Jaques herself told NBC News that she avoided using AI to write the new paper, using LLMs only as a form of inverted inspiration — she types a rough version of what she is trying to say into a model and uses the bland response as motivation to write it herself.

Why it matters

• The study provides peer-reviewed empirical evidence that AI does not merely shift writing tone: it changes content, argumentation and authorial identity in ways users do not consciously detect.
• The "blandification" phenomenon has implications well beyond individual writing: it affects academic evaluation criteria, journalism, institutional communication, and any domain where written expression is central.
• The gap between users' reported satisfaction with AI-assisted texts and their acknowledgement that those texts are less creative and less in their own voice raises serious questions about how humans will evaluate and preserve the authenticity of written expression over time.
• The study opens an urgent research agenda on the impact of LLMs on human values, institutions and modes of thought — an area that, according to Jaques herself, remains largely unexplored.