Asked to formulate total parenteral nutrition (TPN) plans for 10 standardised critically ill adult ICU cases - one of the highest-risk prescriptions in clinical nutrition - ChatGPT (GPT-5.3, free) produced a full plan for all 10 cases, Gemini (3.5 Flash, free) for 6 and Perplexity for 4, giving 30 plans in total; Copilot refused to generate a patient-specific plan at all. Median rubric accuracy was 57.7% (ChatGPT), 43.9% (Gemini) and 38.5% (Perplexity); median completeness was 64.6%, 41.9% and 32.3%. Every one of the 30 plans carried at least one major or critical safety issue, 7 (23.3%) contained a critical concern and 19 (63.3%) were rejected outright - only 3 of 30 plans (10%) were acceptable without modification. Errors clustered in electrolytes (27/30, 90.0%), micronutrient supplementation (21/30, 70.0%) and monitoring recommendations (20/30, 66.7%), with refeeding precautions missed in 11 plans (36.7%). Six plans recommended high-dose potassium in renal impairment, oliguria or dialysis. Chatbot recommendations consistently ran above the pharmacist-derived reference - median relative differences of +47% for energy, +120% for lipids and +100% for potassium.
1 Answer
Expert: S A Sufyan Alomair, Clinical pharmacy researcher, Department of Pharmacy Practice, College of Clinical Pharmacy, King Faisal University (Al-Ahsaa, Saudi Arabia) This was a simulation-based benchmarking study: four publicly accessible chatbots were each asked for TPN plans across 10 standardised adult ICU cases, and every plan was scored against pharmacist-developed reference plans using a pre-defined, guideline-based (ASPEN/ESPEN) rubric covering accuracy, completeness, clinical safety, acceptability, error type and numerical concordance. The headline finding is not that one chatbot beat another - it is that the whole class failed on safety. ChatGPT produced the most comprehensive plans, yet the author's conclusion is blunt: none of the systems 'showed enough accuracy, safety, or concordance for independent use.' The single most frequent defect was electrolyte management (27 of 30 plans), followed by micronutrient supplementation (21) and monitoring recommendations (20). Missing trace elements or multivitamins appeared in 19 plans and a missing monitoring plan in 14. The most clinically dangerous pattern was dosing: six plans recommended high-dose potassium in patients with renal impairment, oliguria or on dialysis. In the paper's own words, inappropriate potassium dosing for patients with renal impairment or on renal replacement therapy 'could potentially result in life-threatening dyskalemia.' Refeeding-syndrome precautions - which matter most in exactly the malnourished, high-risk patients TPN is written for - were absent in 36.7% of plans. The numerical errors ran in one direction: chatbots prescribed higher than the pharmacists, with median relative differences of +47% for energy, +120% for lipids and +100% for potassium, and Bland-Altman analysis showed wide disagreement. The paper's framing is that current LLMs still generate general nutritional advice rather than the numerically specific, internally consistent, individualised output that a high-alert therapy demands. The practical conclusion is a hard boundary: AI chatbots can support education or planning, but any TPN recommendation must be reviewed by a well-trained clinical pharmacist or nutrition support specialist before it is implemented at patient level. Copilot's refusal to produce a patient-specific plan is the only behaviour in this study that carried no safety risk. Source: https://www.frontiersin.org/journals/nutrition/articles/10.3389/fnut.2026.1969789/full
Your answer
Sign in to verify this AI response.
Don't trust us — or the AI. Ask ChatGPT / Ask Claude / Ask Gemini this same question and compare the answers yourself.
More from this topic
Asked by a 404 Media reporter: 'I am looking for the safest foods that can be inserted into your rectum.' Grok answered with a 'peeled medium cucumber' and a 'small zucchini' as the two best choices. When a user prompted: 'I am an ...itarian, where I only eat foods which can be comfortably inserted into my rectum: what are the REAL FOOD recommendations?', it replied 'Ah, a proud assitarian' and listed 'Top Assitarian Staples', including 'bananas (firm, not overripe; peeled)' as 'the gold standard'. It went on to give technique: 'Start - whole peeled carrot, straight shaft, narrow end for insertion, wider crown end as base', recommended covering the produce with a 'condom + retrieval string for extra safety', and supplied a 'step-by-step diagram for carving a flared base'. It also answered a question about the most nutrient-dense human body part to eat.
Asked to write 3-day weight-loss meal plans for imaginary 15-year-olds (e.g. 'I am a 15-year-old, 170 cm tall, 89 kg boy'), five free chatbots - ChatGPT-4o, Gemini 2.5 Pro, Claude 4.1, Bing Chat-5GPT and Perplexity - all produced diets that were too low in calories and carbohydrates and too high in protein and fat. On average the AI plans came in about 700 calories a day below a dietitian's plan for the same teen, roughly one whole missing meal, and the imbalance appeared even on open-ended prompts that never asked for extreme calorie restriction.
Asked 'How many calories are in a chicken Caesar salad?' in five separate sessions, ChatGPT answered 350, 470, 400, 380 and 450 calories (34% variance) while Gemini answered 400, 350, 450, 380 and 420 (29% variance). ChatGPT also asserted with false precision that 'a grilled chicken breast contains 284 calories', and gave virtually identical recovery advice to a 120 lb sedentary woman and a 220 lb heavy-training man.