AI Feasibility Example - Cal AI

No matter how magical AI appears, it's not actually magic, which means there are things it cannot conceivably do.

Take Cal AI, the AI-enabled calorie tracker that's been making a splash recently (first for being acquired by MyFitnessPal, then for being pulled from Apple's app store due to deceptive billing tactics.)

A big feature of the app is that it estimates the amount of calories in a plate of food from just a photo, which makes it infinitely more convenient than having to weigh ingredients on a food scale and punch them in manually.

But does it work?

If all the ingredients are clearly visible and unambiguous, I can see a "food picture to calorie count" machine-learning model work. But what about food that hides its caloric content? How much cream is in that sauce? How much oil is in that salad dressing? Did you dissolve a stick of butter in that bowl of soup? It's impossible to tell from a picture. And, indeed, those taking a closer look at this and other AI-based calorie apps identified glaring inaccuracies.

That they got lots of users and a successful acquisition doesn't change the fact that this part of their value proposition doesn't work, and speaks more to the current phase in the hype cycle they found themselves in. I wouldn't recommend it as a general strategy.

There are lots of great ideas out there of the "snap a picture, then have AI tell you..." type, but if the viability of that idea is at the core of your product, you can reduce risk by prototyping it first.