How LLMs Distort Our Written Language
This presentation examines how Large Language Models systematically alter human writing through empirical evidence from three studies: a controlled human user study, a counterfactual analysis of essay revisions, and an institutional audit of scientific peer reviews. The research reveals that LLMs don't just polish text—they fundamentally reshape semantic meaning, argumentative stance, and authorial voice, creating a collapse toward homogenized, neutral, and impersonal prose. The findings expose critical risks as LLM-mediated writing infiltrates institutional processes, potentially transforming collective norms around creativity, dissent, and epistemic diversity.Script
When you ask a language model to help polish your writing, you might expect grammatical corrections and clarity improvements. But what if the model is systematically rewriting your conclusions, neutralizing your arguments, and erasing your voice?
The researchers conducted three complementary studies to quantify exactly how language models distort human text. They tracked real users writing with AI assistance, compared AI edits against expert human feedback on the same essays, and analyzed thousands of scientific peer reviews to measure institutional-scale effects.
The first major finding reveals something striking about how AI reshapes meaning itself.
Look at what happens when a language model revises a human essay about self-driving cars. The red text shows what the model removed—personal anecdotes, specific concerns, clear conclusions. The green text shows what it added—generic statements, hedged language, impersonal constructions. The model didn't just edit for clarity; it fundamentally rewrote the author's argument, stripping away voice and conviction in favor of neutral, palatable prose.
The contrast is quantifiable and stark. When human experts revise essays, they preserve the author's intended stance and voice while improving clarity. Language models do the opposite—they push essays toward neutrality, strip out personal pronouns and anecdotes, and replace far more vocabulary than any human editor would. Participants who relied heavily on AI assistance reported their essays as significantly less creative and less reflective of their own voice.
The distortion doesn't stop at individual essays. When the researchers analyzed 18,000 peer reviews from a major conference, they found that 21% were Large Language Model-generated. These AI-written reviews systematically deprioritized clarity and relevance while emphasizing reproducibility and scalability. They also assigned scores a full point higher on average. This isn't just about writing style anymore—it's about AI systems quietly reshaping the values and incentives within scientific institutions.
Language models aren't neutral writing assistants—they're active agents that homogenize thought, flatten voice, and shift collective norms. Visit EmergentMind.com to explore this paper further and create your own research video.
