A diabetic developer snapped photos of his meals. Sent them to top AI models. Asked for carb counts. Did it 27,000 times. Not once did he get the same answer twice.
This isn’t a glitch. It’s the norm. David Kernan, writing on his site Diabettech, tested OpenAI’s GPT-5.4, Anthropic’s Claude Sonnet 4.6, and Google’s Gemini 2.5 Pro and 3.1 Pro Preview. Thirteen real meal photos. Each zapped 500-plus times with identical prompts. Results? Wild swings that could wreck blood sugar control.
Take paella. One query to Gemini 2.5 Pro: 45g carbs. Another: 90g. That’s double. Enough to double an insulin dose. Kernan calculated variance as standard deviation divided by mean. Claude scored lowest at 2.4%. GPT hit 8.4%. Gemini 3.1 Pro: 10.3%. Worst offender, Gemini 2.5 Pro: 11%. But those percentages mask real danger. A 11% swing on 50g paella? 5.5g off. For insulin ratios like 1:10, that’s 0.55 units. Over a day, errors compound.
Kernan didn’t stop at variance. He modeled insulin impact. Assuming 1:12 ratio and 100g bolus correction, lifetime risk jumps. Gemini 2.5 Pro showed potential 42.9 extra units dosed. Deadly if low.
Studies Pile On the Evidence
Conferences echo this. At ATTD 2026, researchers reported AI nails simple foods 95% of the time. Apples. Bananas. Easy. Complex dishes like lasagna? ChatGPT right just 43%, per DiaTribe. Overestimations spell trouble. Dietitians erred big (20g+) 3% of cases. ChatGPT: 13%. Claude: 17%. Gemini: 38%. Extra carbs mean extra insulin. Cue hypoglycemia.
PubMed study backs it. Mean absolute error: dietitians 13g ±10g. ChatGPT 20g ±18g. Gemini 28g ±26g. Claude 23g ±21g. Errors higher for Gemini, Claude. Not ChatGPT, barely. But overestimations? AI three to twelve times worse. PubMed.
Conexiant dug into ChatGPT-4o. Simple fruits, veggies: 93-95% agreement with manual counts within 10g. Mean error 3-5g. Composite meals? 43-47% agreement. Errors 14-18g. RMSE hit 26g. ‘Individuals living with T1D should use ChatGPT-4o… with caution,’ they warn. Complex meals unreliable. Risks hypo- or hyperglycemia. Conexiant.
And variance everywhere. Kernan’s core gripe. Even if average right, flipside wrong. No diabetic risks that roulette.
Why It Fails—and What’s Next
LLMs guess. Probabilistic. Temperature settings add noise. Vision models parse blurry phone pics. Portion sizes? Guessed. Ingredients hidden. Outputs drift.
Diabetes Technology Network UK laid it out: ‘Generic LLMs must never be used as autonomous advisory calculators for insulin delivery.’ Backed by Kernan’s data. Open-source iAPS now hooks into OpenAI, Anthropic, Google APIs for food analysis. Bold. Risky.
But fixes brewing. Fine-tuned models on food databases. NIST carb standards. Multi-image prompts. Size references cut errors in Conexiant tests. Still, experts say oversight essential.
X buzzes with doubt. ‘AI carb counting gives different answers each time. Wrong dose = emergency room,’ posted @BillLi_AI April 29. @nonconformie: ‘Throwing raw LLMs at healthcare is reckless.’ @ZambeziSentinel flags trust erosion.
Diabetes hits 500 million worldwide. Type 1 demands precision. AI tempts with ease. Snap a pic. Get carbs. Dose. But inconsistency kills that dream. For now. Developers chase deterministic tools. Regulators eye guardrails. Patients? Manual counting holds. Scale weighs king.
WebProNews is an iEntry Publication