Depending on AI for writing? It quietly distorts what you mean to say

BENGALURU: When researchers at UC Berkeley, UC San Diego, and Google DeepMind asked a group of 100 people to write an essay on whether money leads to happiness, they weren’t just studying writing.

They were studying something more unsettling: what happens to your argument when you hand it to an AI.The answer, according to a new paper published this month, is that it often disappears. The study found that participants who heavily relied on a large language model (LLM) to write their essays were nearly 70% more likely to produce a neutral piece — one that didn’t actually take a side — compared to those who wrote without AI assistance. The people who used AI the most also reported, paradoxically, that the writing felt less creative and less like their own voice. Yet they were equally satisfied with the result.That last detail may be the most telling. “People are satisfied with the results, even though their voice and creativity are diminished,” the researchers write, a dynamic they describe as the central paradox of AI-assisted writing.The paper, titled “How LLMs Distort Our Written Language”, goes well beyond the user study.

The team also ran a large-scale analysis using a pre-existing dataset of 86 student essays on self-driving cars, collected in 2021 before ChatGPT's release. They fed those essays, along with expert human feedback, to three leading AI models, asking each to revise the text in five ways, from a full general rewrite down to grammar corrections only.In every case, the AI-edited essays shifted dramatically in semantic meaning, moving in a strikingly uniform direction across all models and all prompts.

Human revisions, by contrast, were small, varied, and scattered across the embedding space. “Even the instruction to make minimal edits shows large shifts,” the researchers note. AI models pushed essays toward a region of semantic space where no human-written essay had previously existed.One finding stood out as particularly difficult to explain away: when an AI was asked to fix only grammar, it still changed the argument.

In one illustrated example, a student’s conclusion that America “is not ready for self-driving cars” was subtly reframed after a grammar edit to suggest the technology simply needs more efficient implementation first, a meaningfully different claim.The researchers also examined the real-world consequences by analysing 18,000 peer reviews from ICLR 2026, a leading machine learning conference where roughly 21% of reviews were found to be AI-generated. The results were striking. AI reviewers were significantly less likely to flag clarity or relevance as strengths or weaknesses — criteria human reviewers rely on heavily — and instead gravitated toward reproducibility and scalability. AI-written reviews also gave scores, on average, a full point higher than human reviewers.“LLMs have begun to change the very criteria that researchers use when evaluating peer-reviewed scientific research,” the authors write.

The paper suggests a structural explanation for why this keeps happening. AI models are trained with reinforcement learning from human feedback to maximise broadly positive responses, but without any mechanism for modeling what a specific individual actually wants to say. The result, researchers speculate, may be a kind of written clickbait: statistically pleasing, emotionally amplified, analytically polished, and subtly someone else’s opinion.The authors stop short of recommending people stop using AI for writing altogether. Their data suggests that light use, treating the model as an information-seeking tool rather than a ghostwriter, largely preserves the writer’s voice and argument. It's the wholesale handoff that erases them.