Using AI to help you eat healthier
I fed food labels into AI and asked it to suggest improvements. Here's what happened.
Takeaways from this article:
AI isn’t always trustworthy, but it can easily help you eat healthier and improve your nutritional density.
Photograph all the food you consume for one week. Including all nutrition labels and ingredient lists. AI will easily spot healthy and ultra-processed foods.
Ask AI to determine how healthy your diet is. Question the AI multiple times and different ways, to get a sense of whether its judgment and advice are accurate.
Ask the AI help you improve your nutritional health. You can ask it to specifically prescribe how to more closely adhere to a specific diet — Mediterranean, keto, flexitarian, paleo (or the AHEI diet, identified in this study as the one most greatly associated with reaching age 70 in good health).
After Harvard published research in March suggesting an association between eating a healthy, plant-centric (but meat-inclusive) diet in midlife and reaching old age in good health and free of chronic disease, I got curious.
Were my daily eating habits putting me in a position to live a long, healthy, athletic and cognitively sharp life? This is a question many of us are asking nowadays, especially as the negative consequences of eating ultra-processed foods — which in the United States, at least, are everywhere — become brutally clear.
So I began an experiment.
But first, in case you missed it:
During one week in April, I photographed the nutrition labels of every food item I consumed that had a label (see my photo collage above). I fed those 26 photos into one of the more trustworthy AI applications, along with all the other food I consumed that didn’t have labels (e.g., chicken and fish filets; vegetables; berries).
I wanted AI to analyze it all and tell me what I was doing well and not so well. What follows are the results of that experiment — one that anyone can use as a template to critique their own eating habits and routines.
It’s not complicated, if you’re willing to use your smartphone to shoot some food labels, and then feed those labels into an online AI application for analysis.
Step 1: Thoroughly document the food you eat for 7 days
Photograph every raw food, and every packaged food label, that you use to build meals for one week. Skip nothing. Collect those photos into an electronic folder.
That’s it.
Step 2: Upload your weekly food photos to an online AI app
Once you have your food and food label photos in one computer folder, go to your preferred AI app online and drag the photos onto the app’s prompt window, to upload them:
The Big 4 AI apps can each be used for free, but you may consider buying a month’s subscription — they each cost $20/mo — to eliminate restrictions in the free versions on the number of your queries and the amount of information you’re allowed to upload:
Open AI’s ChatGPT. Allows only 10 uploads at a time, limiting its usefulness.
Google’s Gemini. Allows only 10 uploads at a time, but actually uploads the first 10 files of a larger batch, which is helpful.
Anthropic’s Claude1. The free version allows up to 20 files at once per chat. Each file must be no larger than 30MB. (I used the paid version to analyze my diet.)
Perplexity. The free version restricts file uploads to 3 per query, making it the worst of the four listed AIs for analyzing your dietary routines.
Once you’ve uploaded all the evidence of your weekly diet into AI, it’s time to ask it for a detailed analysis. This is called prompting.
Step 3: Prompt AI to analyze your weekly diet
The quality of the prompt really matters. Punctuation and brevity don’t.
What makes a good prompt is being as specific, explicit and comprehensive as possible about what you want it to do and how you want it to organize the information it gives you. Assume it can do almost anything — because sometimes its quality can be stunning — and ask it for everything. The AI may not give everything, but at least by asking, you don’t leave any potentially helpful feedback on the table.
Here’s the relatively brief, modest prompt I initially used for Claude — after several iterations to correct the feedback it was giving me— to get some helpful information:
PROMPT: ”These are food labels of foods I ate during one week. what can you tell me about the quality and quantity of the nutrients of my diet, including amount of processed food, the total grams of protein, the total amount of fiber, the total amounts of all the vitamins and minerals shown on the labels, the total cholesterol, the total saturated fat, the total unsaturated and monosaturated and polyunsaturated fat, and the type of diet regime my eating habits most likely follow.”
This prompt produced a nutritional analysis of my weekly protein, fiber, fat, sodium, cholesterol, carbs and vitamins and minerals, as well as nutritional considerations and “personalized recommendations,” culminating with this eyebrow-raising conclusion:
Everyone likes being flattered. But not by a machine that’s pretending to be serious. I eat healthy, but not enough to warrant this level of artificial inanity.
So I asked Claude for a reality check, and it responded as I suspected it would:
Step 4: Ask the AI if it’s being accurate and comprehensive
What happened next is slightly frustrating but instructive — and also worth the effort of holding AI to account. Because, held to account, AI can be a huge help to improving nutritional strength as we age.
Claude proceeded to give me another comprehensive nutritional assessment that, while not overtly glowing, still made me wonder. When I asked it if it was being too complimentary, Claude again said it was. So I prompted it, again, to give me a fair and accurate reading of my weekly diet.
This time, however, Claude’s analysis swung too far into the negative. So I prompted it yet again:
It’s kind of funny, kind of sad. But that’s the state of consumer-level AI for now (but probably not for long). The five or so prompt iterations and challenges I went through are, I think, absolutely worth the few minutes they take (AI answers come back in seconds). Here’s what Claude finally gave me as a concluding analysis of my weekly dietary intake, which I found helpful:
This reasonably accurate assessment came at the end of a long and quite detailed analysis of my diet that included recommendations for improvement. Which is why the fifth and final step is important.
Step 5: Ask AI how to strengthen your nutritional health
When AI doesn’t give you what you’re looking for, as it again, with as much specificity and detail as possible.
At the top of this post, I mentioned the AHEI diet, which Harvard public health scientists recently identified as the one with the greatest association to healthy aging. I asked Claude what I should do to make my diet align more closely to the AHEI diet. This screenshot is a fraction of the specific and useful advice it gave me:
What do you think? Let me know in a comment, please. If you want to see Claude’s full response to my questions about becoming nutritionally stronger, I’ll send it to you.
I’ve found that Claude often provides the most useful and comprehensive information, but like all AI apps is susceptible to exaggerations and outright misstatements.
I'm not sure how useful it is. AI just tells you what it thinks you want to hear it seems. Even when you ask it to be more realistic, it is still just telling you a version of what it thinks you want
what a great idea Paul to solicit AI assistance...thank you. My partner has a helpful guide as well, for what it's worth: try only eating what is recognizable for what it actually is...