The One Thing AI Vision Still Gets Wrong

I ran a few photos through AI just to see where it breaks. Some of the results made me trust it more. One of them made me trust it a lot less.

First, a whiteboard covered in messy math, my own head blocking half the words. AI figured out I was teaching convolutional neural networks anyway, even with my skull sitting directly on top of the word convolutional. Check mark.

Then a handwritten letter in cursive. Most people under 40 can't read cursive anymore. AI transcribed it clean. Try it yourself sometime, you'll probably lose.

Then a restaurant receipt. I gave it my three items and asked for my share of the bill. It did the math correctly. (Still double check anything with actual money attached, but it nailed it.)

Then a photo of a weird human sized hamster wheel treadmill. I asked for a sales ad. It wrote a good one, because the object is visually distinct enough that there's no mistaking what it is.

So at this point you'd think AI vision just sees everything. Then I showed it a photo of the machines at my gym and asked what they were. It gave me a confident, detailed, completely wrong answer.

Why the Gym Photo Broke It

The difference isn't complexity. Handwriting is arguably harder to parse than a leg press machine. The difference is visual distinctiveness.

A cursive letter still has unique shapes for every word. A treadmill shaped like a giant wheel doesn't look like anything else on earth. But gym machines, through a slightly blurry photo, all kind of look like the same tangle of metal and cables. AI doesn't know what it doesn't know here. It just picks the closest match and says it with total confidence.

AI doesn't fail quietly, it fails confidently, which is the actually dangerous part.

The Rule for Using AI's Eyes

Use it for the gist, not the fine print. Summarizing a whiteboard, transcribing a recipe card, digitizing a stack of brainstorm notes, all great uses. It reads the room fine.

But when the details actually matter, when two things look almost identical and only differ in some small feature, don't trust the first confident answer. Check it yourself, or ask a followup question to see if the answer holds up.

Next time you hand AI a photo, ask yourself if you're looking for the big picture or the fine print. That answer tells you how much to trust what comes back.

#AITools #AIWorkflows #PromptEngineering #AIVision