More Words, Less Thinking: The Paradoxical Homogenization of AI-Assisted Writing
AI makes writing more eloquent — and less original.
Table of Contents
Table of Contents
AI makes writing more eloquent but less original. A Georgetown study of 373,000 college admissions essays found that after ChatGPT became widely available, essays used richer vocabulary but expressed narrower, more similar ideas. The author calls this paradoxical homogenization: the illusion of originality while our thoughts quietly converge.
The Georgetown Study: 373,000 Essays, Before and After ChatGPT
Researchers at Georgetown University analyzed a massive corpus of college admissions essays — 373,000 in total — comparing submissions written before and after ChatGPT. The dataset was large enough to detect subtle shifts in both language and meaning. The pattern was clear: fluency went up, variety went down.
Post-ChatGPT essays showed measurable gains in lexical sophistication. Words became more polished, sentences more fluid, and overall prose more confident. But underneath that surface improvement, the ideas themselves were becoming less diverse.
The Paradox: Richer Vocabulary, Narrower Ideas
That is paradoxical homogenization in practice. Students now have a tool that can help them express themselves more fluently, yet the very fluency it produces tends to smooth away the rough edges that make an essay memorable.
LLMs are trained to predict the most probable next token. Statistically, that means they gravitate toward common phrasings, safe examples, and familiar conclusions. When many students use the same assistant, the outputs naturally cluster around the same linguistic center of gravity. The result is writing that sounds better but thinks less broadly.
Why Admissions Essays Matter as a Test of Originality
College admissions essays are a revealing benchmark because they explicitly reward personal perspective. Admissions officers are not just grading grammar; they are looking for insight, voice, and evidence of independent thinking.
If a tool can make an essay look more impressive while diluting the unique point of view, it undermines the signal the essay is supposed to send. The essay becomes harder to trust as a window into the applicant’s mind.
How LLMs Steer Toward Statistically Probable Answers
Large language models optimize for likelihood. They learn from vast corpora of human text and generate responses that fit the patterns they have seen most often. That makes them excellent at producing conventional, middle-of-the-road content.
The problem is that original ideas are often unlikely. They sit at the tail of the distribution, not the center. A model designed to maximize probability will therefore tend to avoid them unless it is deliberately prompted to take risks.
The Illusion of Originality: Fluent Prose Is Not Original Thought
A well-written paragraph can create a powerful impression of competence and intelligence. But fluency and originality are not the same thing. An essay can be grammatically perfect, stylistically polished, and conceptually generic at the same time.
The danger is that readers — including teachers, admissions officers, and recruiters — may start to mistake style for substance. As AI-assisted writing becomes the default, we will need better ways to distinguish genuine insight from surface-level eloquence.
What This Means for Education and Assessment
For educators, this changes what a writing assignment can measure. Traditional writing assignments may become less reliable as measures of student learning if students can generate plausible prose with minimal effort. The assessment needs to shift toward process, reflection, and the kind of reasoning that cannot be outsourced.
One path forward is to value the unlikely idea: the observation that only emerges from a particular student’s experience, the argument that runs against the obvious consensus, the question that the model would not have thought to ask.
Conclusion: The Value of Unlikely Ideas
The Georgetown study does not claim that AI is making students worse. It shows that better tools can make us more articulate while making us more alike. The harder task is to preserve — and teach — the habits of mind that produce genuinely original thought.
When fluent AI prose becomes the default, the scarcest and most valuable skill may be the willingness to pursue an idea that is not the most probable one.
