Large language models to write scientific manuscripts: to be considered but not trusted

Large language models (LLMs) are rapidly permeating scientific writing. Using ChatGPT 4o and real-world data from the Swedish Heart Failure Registry, we explored both the promise and the pitfalls of LLM-assisted manuscript drafting. We demonstrate that, with strict human oversight, LLMs may accelerate text generation but still hallucinate, frequently misinterpret effect directions, and are often unable to access other manuscripts for direct references. Until these limitations are addressed, LLMs should complement, rather than replace, domain experts in scientific publishing.