Seonglae ChoWhen LLMs are used as writing-assistance tools, they do more than introduce superficial stylistic changes—they can systematically distort the intended meaning itself. This is quantified through three independent studies (an RCT user study, counterfactual editing analysis, and an ICLR peer-review analysis). In a randomized controlled trial with 100 participants, people wrote an essay on “Does money lead to happiness?”, comparing a condition with access to an LLM (gpt-4o-mini) against a control condition. The authors visualized semantic shifts using MiniLM-L6-v2 sentence embeddings with t-SNE/PCA, and measured differences in lexical distributions via Jensen–Shannon Divergence (JSD). They also quantified emotional and psychological language shifts using the NRC Emotion Lexicon and LIWC. Across edits, LLM-assisted revisions produced large semantic moves in a consistent direction in embedding space—contrasting with human edits, which tend to be smaller and more multi-directional.
Even when the prompt asks only for grammar correction, the LLM tends to change the essay’s conclusion. In the heavy-LLM-use group, the share of essays taking a neutral stance increased by 68.9% (). Participants also reported significantly lower creativity (, ) and a loss of personal voice (, ). On ArgRewrite-v2, the LLM’s JSD was nearly 3× the human baseline (0.2–0.3), reaching an average of 0.476; pronoun usage decreased by 40–60% while adjective usage increased by 57–90%. Positive sentiment increased by 37–54% and negative sentiment by 24–38%, indicating a general amplification of affective language. In an analysis of 18,000 ICLR 2026 peer reviews, LLM-generated reviews had scores about 10% higher on average than human reviews (4.43 vs 4.13); they were 32% less likely to cite clarity as a strength (, ), and 136% more likely to mention reproducibility.
Analysing the impact of ChatGPT in research
For the same paper, AI-assisted reviews gave an average score 0.14 points higher than human reviews.