Real data gradually moves away from the training data. The model doesn't break by itself. The world changes: new scams, slang, prices, and the old training no longer fits them.

Why did the scam filter miss a new scheme?

Examples like it weren't in the training data. A model knows only what was in its data. It will catch the new scheme only after it is retrained on fresh examples.

Why press “not spam” or “bad answer”?

Feedback becomes new labels for testing and training. Feedback is collected and used to find mistakes and retrain the model. The effect shows up by the next update, not instantly.

AI in real life: models get old

Lesson 19 of 23 in the course “AI basics”. Free, with animations and a quiz.

A model is a snapshot of the past

Example: a banking app runs a filter that catches scam text messages. It was trained on old messages labeled “scam” or “normal”. A label is the correct answer for an example, set by a person. The filter learned only what was in that data. If a label was wrong, it repeats the mistake. If some scheme wasn't in the data, it knows nothing about it.

The world changes, quality drops

A year goes by. Scammers come up with something new: “side job, like videos for cash”. Messages like that weren't in the training data, so the filter lets them through. This is data drift: over time, real data differs more and more from the data the model was trained on. The model didn't break by itself; the world moved on. Slang, prices and trends go out of date the same way, and the model's answers with them.

Watch it and update it

That's why a model inside a product is watched. This is called monitoring: the team tracks how many scams slipped through and how many complaints came in. Quality dropped? They collect fresh labeled examples and retrain the model, meaning they train it again, this time on new data. On the chart, that's the jump back up. Chatbots work similarly: new versions come out because the old ones' knowledge goes stale.

Human in the loop and the feedback button

Human in the loop is a setup where a real person reviews the AI's important or disputed decisions. For example, blocking a card or turning down a loan. Your “not spam” or “bad answer” button is part of the loop too. That feedback becomes new labels for testing and training. One click doesn't change the model right away, but by the next update the feedback helps fix its mistakes.

Responsible AI

Responsible AI rests on three things. Fairness: if the data has few people with a certain accent or skin tone, the model will work worse for exactly those people. Privacy: users' personal data is protected and not handed out. Transparency: people know they're talking to a bot and can find out why it decided what it did. Key rule: if an AI decided something important about you, ask for a human to review it.

Check yourself

What is data drift?
Correct answer: Real data gradually moves away from the training data. The model doesn't break by itself. The world changes: new scams, slang, prices, and the old training no longer fits them.
Why did the scam filter miss a new scheme?
Correct answer: Examples like it weren't in the training data. A model knows only what was in its data. It will catch the new scheme only after it is retrained on fresh examples.
Why press “not spam” or “bad answer”?
Correct answer: Feedback becomes new labels for testing and training. Feedback is collected and used to find mistakes and retrain the model. The effect shows up by the next update, not instantly.

← Safety: your data and prompt injection Where AI really helps →