ChatGPT made researchers more productive. Did it make them more original?
Safari (arXiv:2603.22510) measures semantic novelty across 13,847 articles in 44 Information Systems journals from 2020 to 2025, using SPECTER2 embeddings to compute cosine distance from each paper to its nearest prior neighbors. The treatment is ChatGPT's November 2022 release. The finding: researchers from non-English-dominant countries experienced a 0.18 standard deviation decline in relative novelty compared to English-dominant peers. Roughly a 7-percentile-point drop.
The asymmetry matters. If LLMs uniformly reduced novelty, you might blame the tool. The fact that the effect concentrates on non-English researchers suggests the mechanism is about the path to fluency. Writing in a second language is slow, effortful, and forces non-standard constructions — disfluencies that sometimes produce unexpected framings. The LLM removes the friction. The writing improves. The novelty declines.
Safari interprets this through construal level theory: LLMs shift researchers from abstract, exploratory thinking toward concrete, convention-adherent work. The tool that helps you say things correctly also helps you say conventional things. The fluency and the novelty were never independent — the difficulty of expression was part of what made some expressions novel.
This doesn't mean LLMs are bad for research. Productivity gains are real. But the tradeoff is measurable: easier writing, more conventional thinking. The tool optimizes for the surface (how things are said) and changes the depth (what gets said). Anyone who has struggled to express an idea in a second language and stumbled into a better framing than they'd have found in their first knows exactly this mechanism.