Skip to content
Train Your Own ModelLesson 9 of 12

9 · Where trained models show up in real life

The same kind of model you just trained — examples in, pattern out — is running quietly all around you. Once you can spot it, you'll see it everywhere. Every one of these is "features in, label out," scaled up:

Where you've seen itWhat it's classifyingFeatures → Label
Phone unlocking by your faceIs this the owner?camera pixels → "match / no match"
Photos app finding "all dogs"What's in each picture?image pixels → object labels
Email spam filterJunk or real?the email's words → "spam / not spam"
Voice assistant hearing "hey ___"Did someone say the wake word?microphone sound → "wake word / not"
Video captions / "what song is this?"Which words? Which song?audio → text or a song label
A bank flagging a weird chargeIs this purchase suspicious?spending pattern → "ok / check this"

Some of these do real good — captions help people who are deaf or hard of hearing; medical models can help doctors spot problems earlier. And every one carries the lessons you just learned:

  • Each was trained on examples, so each can be wrong, and each can pick up bias from its data. A face-unlock that saw few darker-skinned faces, or a voice assistant trained mostly on one accent, will work worse for the people it didn't see enough of.
  • The stakes are higher than your thumbs-up demo. When a model helps decide things like loans, jobs, or who gets flagged by police, a biased model can hurt real people. That's why grown-ups push hard for fairness and accountability in these systems.

Knowing how these are built makes you a sharper, safer user of them. When an app's AI gets something wrong, you won't think it's broken or magic — you'll think, "its training data probably didn't cover this case," which is usually exactly right.

Think about it. Pick one model from the table. What examples do you think it was trained on — and who might it work worse for if those examples left some people out?

Sources

Sign in to track your progress