Abstract The expanding use of Large Language Models (LLMs) in scientific writing raises critical concerns regarding research accountability, integrity, and transparency. This study aimed to develop and content-validate the CAREFUL-AI framework and apply it to evaluate the scientific quality of AI-generated research manuscripts. A three-round Delphi process with 10 multidisciplinary experts was conducted to develop the CAREFUL-AI framework. Content validity was assessed using item-level and scale-level content validity indices. Using standardized, locked prompts, 150 complete scientific research manuscripts across five study designs were generated by six freely accessible LLM platforms ChatGPT (GPT-5.2), Claude (Claude 4.5), Gemini (Gemini 3 Flash), Grok (Grok 4.1), DeepSeek (DeepSeek-V3.2), and Meta AI (Llama 4 Maverick). Manuscripts were evaluated by the respective prompt administrators using the CAREFUL-AI framework. The framework demonstrated strong content validity (I-CVI: 0.83–1.00; S-CVI/Ave: 0.92). Overall, 38.0% of manuscripts were rated high quality, 52.7% moderate quality, and 9.3% low quality. Claude generated the highest proportion of high-quality manuscripts, while Meta AI produced all low-quality outputs. Reproducibility and uncertainty handling consistently received the lowest domain scores across models. The CAREFUL-AI framework provides researchers with a structured tool to critically appraise AI-generated manuscripts, helping safeguard methodological rigor, transparency, and evidence reliability in scientific writing. Substantial variability exists in the quality of AI-generated manuscripts. Although Claude demonstrated comparatively stronger performance, persistent deficiencies in key integrity domains indicate that AI-assisted manuscript generation requires robust human oversight. CAREFUL-AI provides a content-validated framework to support ethical governance and editorial oversight of AI-assisted research.
Santhosh et al. (Thu,) studied this question.