An AI wrote their diet plan after cancer surgery. Two in three reached their calorie target, against two in five on the usual handout, and the AI may not be the reason

Going home after stomach or bowel cancer surgery often comes with a short leaflet: small meals, soft foods, enough fluid, some protein at each meal. It is sound advice, and it does not say how much protein, or what soft food means for this person after this operation. A team at Tongji Hospital in Wuhan tested whether a large language model could fill that gap. Their randomised trial was published in Supportive Care in Cancer on 29 October.

They screened 187 people having planned keyhole surgery for gastric or colorectal cancer and randomised 90, half to each arm, average age about 61. In the AI arm a research nurse typed each patient's details into a set prompt: operation, weight, blood results, what they could tolerate, other conditions. DeepSeek v3.2 wrote a week-by-week plan with calorie, protein and fluid targets and familiar foods such as congee with minced pork and steamed fish. A dietitian checked each plan for safety in 3 to 5 minutes, without changing its substance, and it was printed. The other arm got the hospital's usual one-page advice and talk from the surgeon or ward nurse.

Thirty days after discharge, judged from three days of food records, 30 of 45 patients in the AI arm (66.7%) reached at least 90% of their calorie target, against 18 of 45 (40.0%) on usual advice. For protein, at least 1.2 g per kg of body weight a day, it was 62.2% against 35.6%. Fatigue, nausea, appetite loss and gut symptom scores were lower in the AI arm too. Dietitians who did not know the groups rated the plans higher on completeness, tailoring and practicality, but not significantly higher on accuracy or safety. Staff time per patient was 8.2 minutes against 12.5.

Now the limits, several of them stated plainly by the authors. Weight change (-2.1% against -2.5%), albumin, haemoglobin, complications (20.0% against 22.2%) and nutrition-related readmissions did not differ, though 30 days and 90 people could easily miss such changes. The AI arm got tailoring, written targets, nearly three times the words, and a dietitian's check, all at once, so the authors say the trial does not show AI beats similar tailored advice written by a person. 93.3% of AI plans carried numeric targets against 51.1% of the usual advice, and that may be the active ingredient. Patients knew which arm they were in and kept their own food records, the methods describe no correction for the many symptom scales tested, and this was one hospital with patients who were fairly well nourished. No competing interests were declared; funding came from a Chinese clinical nutrition research foundation.