What you should not trust AI to do with your diet

Not very long ago, Google’s new AI search feature was in the news for suggesting that people put glue on pizza to stop the cheese sliding off. It also recommended eating one small rock a day, presumably for anyone who felt their diet was lacking gravel.

Google later explained that the system had picked up joke and satirical material online and presented it as useful advice. Changes were made, and AI has improved considerably since the days when lunch came with PVA.

But improved does not mean infallible.

Nobody is likely to drizzle glue over a Margherita because Google said so. A calorie total that looks perfectly reasonable is much easier to believe.

The warning label beside the prompt

I recently shared the full prompt I use to turn AI into a weight-loss companion.

I still think it can be remarkably useful. I use it every day to log food, calculate rough calories, plan meals, spot patterns and talk me down when the scale has behaved like an absolute cow.

But a good prompt does not make AI medically qualified, incapable of mistakes or responsible for the decisions I make.

This article belongs beside that prompt as its warning label.

It is not a reason to stop using AI. It is a reminder to keep my own judgement switched on while I do.

The 800-calorie tuna salad

Recently, I told ChatGPT I had eaten my usual tuna salad. It normally comes to about 350 calories.

ChatGPT added nearly 800.

My total for the day was suddenly approaching 1,500 calories, with half the day and dinner still ahead of me. For one horrible moment, I thought my innocent little tuna salad had somehow acquired the calorific properties of a family lasagne.

I was already wondering how many times I would need to run around the block before I was legally permitted to eat again.

I knew the number was ridiculous, so I gave ChatGPT the ingredients and quantities and asked it to calculate the meal properly.

It got it wrong again.

This time, I had said 80g of rice and it assumed I meant 80g weighed dry. I meant 80g cooked rice, as I always had in that salad.

That small but important detail had disappeared from its immediate memory. Instead of asking me whether the rice was cooked or dry, it quietly chose an answer and carried on calculating with complete confidence.

Sometimes AI gets the arithmetic wrong.

Sometimes it performs perfectly good arithmetic on an imaginary version of your lunch.

It can only calculate what you tell it

AI cannot see what happened in my kitchen.

It does not know about the oil I used to fry the onions unless I mention it. It cannot see the butter, mayonnaise, sauce or dressing. It does not know whether the tuna was packed in oil or brine, whether the chicken was weighed raw or cooked, or whether my tablespoon was level or looked like a small geological formation.

If I leave something out, AI will usually leave it out too.

That means a calorie total can look beautifully precise while being much lower than what I actually ate. The oil still contains calories even if it never made it into the conversation.

The opposite can happen when something is unclear. AI might assume dry rice instead of cooked rice, use the wrong product, count an ingredient twice or attach the full-day total to one meal.

The most dangerous part is not always that it gets something wrong. It is that it can be wrong in such calm, convincing sentences.

If a figure looks surprisingly high or low, I should not immediately change what I eat next. I need to ask it to show every ingredient, quantity and assumption it used.

The maths is not the only thing to question

AI can calculate a calorie target for me, but that target is still an estimate. It does not become a biological fact because it arrived in a tidy table.

If a suggested target leaves me constantly hungry, exhausted or thinking about food from morning until night, I need to pay attention to that instead of blindly obeying the number.

AI can also predict where I might weigh by Christmas if I continue losing at the same average rate. That can be encouraging, but it is a projection, not a promise. Bodies do not lose weight in neat weekly instalments, however much I would appreciate the administrative efficiency.

It can create a meal plan, but only from the life I describe. If I have not told it my budget, allergies, foods I hate, available cooking time or preferred way of eating, the plan may be completely unsuitable while still looking impressively organised.

It can notice possible patterns in my weight, appetite, digestion, sleep or menstrual cycle. Those patterns may be helpful clues. They are not diagnoses.

And I would not trust AI to decide that I should skip dinner, slash my calories or exercise excessively because I ate more than planned. An incorrect food total should not be allowed to turn into a real punishment.

There is a firmer boundary too. AI should not diagnose symptoms, tell me to change medication or replace a doctor, registered dietitian, urgent medical help or eating-disorder treatment.

It can help me understand information and work out which questions to ask. It cannot become the qualified person I should be asking.

Make it show its workings

I still use AI to calculate meals. It saves time and, most of the time, it is enormously helpful.

For a new meal, I give it the brands, quantities and preparation details that matter. I say whether rice, pasta and meat were weighed raw or cooked, whether canned foods were drained and whether oil, butter, sauces and dressings were included.

Then I ask it to list each ingredient separately, followed by the meal total and the full-day total.

For a familiar meal, it is often safer to say:

My usual tuna salad is 350 calories. Add it as 350 unless I tell you that something has changed.

I do not assume that a tidy-looking number has descended from the heavens on a stone tablet.

I check anything that looks odd before I cancel dinner, add exercise or rearrange the rest of my day around it.

Even with those risks, talking to a chatbot about my daily calories still feels infinitely nicer than entering them into a dead-eyed calorie app that does not talk back.

AI can encourage me, help me think and make the whole business feel less lonely.

I just need to remember that friendly is not the same as flawless.