Proofreading with an LLM: it edits what was right

Testing Claude Sonnet and Haiku revealed inconsistencies in proofreading and humanizing texts across five languages. Without specific rules, Sonnet rewrote correct content, while Haiku failed to provide corrected text. In humanization tasks, both models removed the word usually, altering meanings, and Sonnet invented facts. However, implementing specialized rules significantly improved performance, with Sonnet passing all proofreading cases and Haiku passing five, demonstrating that bare requests are unreliable for precise linguistic tasks.