The hardest UX problem in AI food logging is not the camera

I have been thinking about AI food logging from the product side, and the camera is not the hard part.

The hard part is trust.

A photo can usually get you into the right neighborhood. It can recognize rice, chicken, sauce, coffee, fruit, or whatever else is on the plate. But the product still has to handle the messy parts:

• portion size is often uncertain
• homemade meals have hidden ingredients
• restaurant food varies wildly
• the user may only want a fast estimate, not a research project
• forcing too many corrections kills the habit

So the UX I like is not "AI knows everything." It is closer to this:

1. Make the first guess fast.
2. Show the assumption clearly.
3. Make the most likely correction easy.
4. Let the user move on.

For example, if the app sees chicken and rice, the useful flow is not a giant form. It is something like:

• chicken breast or thigh?
• white rice or fried rice?
• small, medium, or large portion?
• save this as a frequent meal?

That keeps the product honest. The user still gets speed, but the UI does not pretend a photo can know every gram, oil, sauce, or brand.

This also changes how I think about "accuracy" in AI consumer apps. Sometimes the better product is not the one that claims the model is perfect. It is the one that makes uncertainty cheap to fix.

That is the angle I am exploring with MetricSync, an iPhone app for quick AI food logging from photos: https://metricsync.download

No medical advice, no magic claims. Just trying to make the common "log this meal quickly" flow feel less annoying.

If you are building AI UX, I think this pattern applies in a lot of places: let the model draft, show the assumption, make correction lightweight, then get out of the way.