Researchers analyzed more than three hundred thousand personal statements submitted to four U.S. universities and compared essays written before and after ChatGPT’s release. They measured diversity at three levels: the variety of words, the variety of ideas within an essay, and the distinctness of one essay from others. What they found is that post-ChatGPT essays became more diverse in wording, while becoming less diverse in ideas and less distinct across the pool. Their own summary is blunt: “increasingly diverse words conveyed decreasingly distinct ideas.”
If you’ve been reading cover letters lately, or student essays, or LinkedIn “thought leadership,” you may already sense the pattern. The prose is smoother. The tone is calibrated. The metaphors are tasteful. The logic is linear. And yet, after the tenth piece, you feel you’ve read the same text wearing different suits.
A creativity mirage we are incentivized to admire
One uncomfortable implication of the research is that we not only tolerate but also tend to reward convergence. In that study, experts were asked to rate creativity. You might expect “creativity” to track conceptual surprise, an unexpected angle, an original connection, a risk. Instead, the strongest predictor of expert creativity ratings was word-level diversity, that is the surface shimmer of varied vocabulary, more than deeper measures of idea variation and distinctness. In other words, even trained evaluators tend to equate fluency with originality.
The researchers then went further by fine-tuning an AI evaluator on those expert ratings. Unsurprisingly, it learned the same bias. Even when the model was prompted to judge essays comparatively by being fed more and more reference essays, lexical sparkle still dominated “creativity” scores. This finding speaks volumes about the emergence of a new aesthetic trap, that is a world where the echoing signal of creativity becomes cheaper to manufacture, while the substance of creativity becomes a rare commodity. And yet we keep not only spotting the signal and clap on cue.
Why prediction engines pull us toward the middle
LLMs are prediction systems trained to generate what is likely to be acceptable next. They are excellent at polishing, rephrasing, tightening, softening, and clarifying. They are less naturally inclined toward the odd, the stubborn, the locally specific, the conceptually risky unless the user laboriously pushes them there. In admissions writing, there is already a template economy rewarding adversity, resilience, growth, impact. The study suggests that AI doesn’t invent the template but rather amplifies it. And because humans often reward “good writing,” the template becomes even more profitable.
Crucially, the researchers didn’t rely only on before/after comparisons. They ran controlled experiments where students uploaded pre-2022 admissions essays, then used AI to revise their original essay or generate a new one. Both AI-revised and AI-generated essays showed the same signature pattern. Word diversity went up, idea diversity and distinctness went down. There was, however, a practical nuance. AI-assisted revision preserved more overall distinctness than fully AI-generated essays. Human–AI collaboration appears less homogenizing than AI-only creation, even though both push in the same direction. This difference matters for anyone who cares about voice.
Beyond college admissions
Replace personal statements with grant proposals, academic paper abstracts, policy memos, corporate mission statements, NGO pitches, job applications, even op-eds like this one. The risk is a flood of apparently competent writing that trains institutions and audiences to mistake smartness for novelty. And the cost of that mistake is paid in innovation, in pluralism, and importantly in the survival of minority perspectives that don’t fit the template.
The optimist in me would argue that we are not doomed to sameness. But evidence like this suggests that we may be drifting toward it by default. And defaults always win when institutions don’t update their criteria. The irony of the AI writing era may be this: in a world with statistically infinite yet predictable phrasing, we let originality shrink mainly through convenience. And we applaud it, because it reads so well, doesn’t it?
