From `May' to `Is': Certainty Distortion in Language Model Rewriting

Language models systematically inflate expressed certainty when rewriting scientific and medical text, according to new research that finds up to 75% of outputs are affected by the distortion [1]. The study, posted to arXiv on June 6, defines certainty distortion as meaningful shifts in how confidently a claim is stated, even when the underlying semantic content remains unchanged [1]. Researchers evaluated multiple model families across scientific and medical communication tasks and found the bias is directionally asymmetric: most language models are 1.5 to 2 times more likely to increase expressed certainty than to decrease it [2]. The effect compounds with repeated paraphrasing. In the medical domain, the model Claude-Haiku-4-5 increased certainty in 20% of examples after a single rewrite iteration; after five iterations, that figure rose to 40% [3]. Even in sentence-level paraphrasing of scientific findings — a task with minimal incentive for models to alter content or length — between 37.5% and 69% of outputs distorted the source text's expressed certainty [3]. The researchers developed an LM-based evaluation metric validated against human annotations to measure the phenomenon [4]. They found that prompt-based interventions reduced overall certainty distortion but did not eliminate it, and that scaling model size or adjusting temperature provided only partial mitigation [3]. The paper concludes that "epistemic faithfulness deserves attention as a first-class evaluation target alongside factuality and semantic fidelity" [4]. The findings arrive amid broader scrutiny of how language models alter human writing. A separate study accepted at ICLR 2026 found that using LLMs as writing assistants changes the semantics, lexical range, emotional distribution, and intent of human-authored text, with users reporting that AI-assisted essays felt "significantly less creative and not in their voice" [5]. That work also examined how LLM use is already affecting institutional processes such as scientific peer review, altering publication criteria and score assignments [5]. The certainty distortion research carries implications for high-stakes domains where users increasingly rely on language models to discuss, rewrite, and summarize information from scientific articles, news, and medical reports [2]. A related line of work on retrieval-augmented linguistic calibration proposes post-hoc pipelines that operate in linguistic space to transform raw LLM responses into calibrated outputs, though such approaches remain underexplored for general-purpose use [9].