Caught a machine learning model slacking off at a job site demo

Last month I sat through a vendor demo for an AI inspection tool at a contractor yard in Phoenix. The thing was supposed to flag safety issues in photos of fences and gates. Worked great on their clean sample shots, looked like magic. Then I handed them a photo from my actual job site, dusty, crooked post, harsh noon sun. It missed a broken latch completely. They ran it again, still missed it. Took them 20 minutes to realize the training data was all tidy suburban stuff, nothing like the real field. I get that AI needs good data, but if it can't handle dirt and bad angles, it's just a toy for the office. Anyone else tested these tools with messy real-world photos and watched them fall apart?

2 comments

2 Comments

knight.felix1mo ago

That 20 minutes of fiddling in Phoenix says it all, right? Their whole pitch was basically "trust us," but the second real dirt showed up, the whole thing turned into a $50k paperweight.

hannah4001mo ago

Phoenix is just one city though, and that demo was on a prototype unit. I've seen way worse from companies that actually shipped stuff, at least they were upfront about it being early. Plus, who's really running that thing in the field for 8 hours straight anyway? Most of the time it's just sitting in a truck between jobs. I get being annoyed at the hype, but calling a $50k paperweight after one bad showing feels like a stretch. There's a difference between "needs more work" and "literally useless," and I think people are jumping straight to the second one.