A new framework called GIRISH (Goal, Input, Role, Iterative Refinement, Safety Verification, and Human Accountability) is proposed for structured, human-accountable artificial intelligence (AI) in neonatal intensive care. The framework addresses the growing use of AI in healthcare, including predictive machine-learning models, foundation models and large language models (LLMs), and agentic systems. AI is particularly relevant in low- and middle-income countries (LMICs) facing shortages of trained healthcare professionals and resource limitations. In neonatology, AI is explored for risk stratification, physiologic monitoring, early identification of clinical deterioration, telehealth support, and decision-making assistance. The article cites early experience in India where ChatGPT has been used to provide patient-specific answers to parents’ questions in pediatric intensive care with clinician oversight. A clinical scenario illustrates the risk: a 28-week, 950 g infant on day of life (DOL) 7 deteriorates, with FiO2 rising from 0.38 to 0.55 over six hours, CRP at 48 mg/L, white cell count 18.4 × 10⁹/L with 74% neutrophils, and temperature 37.9°C after seven days of ampicillin. A clinician typing a vague query about late-onset sepsis receives a plausible but potentially wrong answer, highlighting the difficulty of distinguishing correct from incorrect AI responses without a structured approach. Surveys confirm neonatologists already use AI chatbots such as ChatGPT, Gemini, Claude, and