Back to topics

How can an AI’s training examples shape its mistakes?

Imagine a picture classifier trained on clear daylight photographs, then tested on blurry evening images. Familiar examples may not prepare it for the new conditions. Training coverage and the test set become part of the explanation of its errors. This simplified classification example does not explain every kind of generative AI behavior.

3–7 min

Questions you could explore

  • What examples would reveal whether the system learned the intended distinction?
  • How could a misleading label create a different kind of error?

Start reading

Ready to try your ideas aloud?

Your topic carries over. Record up to 2½ minutes, depending on your plan.

Computers and AI

Why can a damaged QR code still work?

A scratch or stain does not always stop a QR code from being read. Its stored information includes redundancy, but that does not mean any part of the image can be covered without consequence.

3–5 min
Computers and AI

How does a website name become an internet address?

You remember a website by its name, while the network needs an address for the connection. DNS supplies that translation, sometimes from a saved answer and sometimes by asking other servers.

4–5 min
Computers and AI

Same emoji, different picture

The same emoji can arrive with a different expression or style from the one you saw when sending it. Unicode identifies characters and sequences, while software platforms choose artwork and decide what their versions support.

2–3 min
How can an AI’s training examples shape its mistakes? | Speech Topic | speaking.app