Rocket holds up a slip sentence from the lab. "Raven logs a comet. That never happened, but it sounds right."
Raven reads another. "Rocket spots a star. That one is copied straight from the log."
"So the predictor makes new sentences out of old patterns," Rocket says.
Nova dims her lights to think. "Notice what it cannot do," she says.
"It cannot add a word it never saw, and it never checks the sky."
"It only knows which word tends to follow which."
Raven taps the cups. "So it is a pattern machine, not a stargazer."
Generative models make outputs that approximate the patterns in their training data.
That is why their output can be new and still feel familiar. It mixes pieces in ways the patterns allow.
The paper predictor shows the idea at a tiny size. Real models learn from tremendous amounts of data.
The AI4K12 Five Big Ideas reminds us that AI does not think the way a person does.
Real generative AI is a tool that people build. It produces content from patterns, and it can be wrong.
Nova can wonder and joke in our story because she is a character. Real AI tools do not feel or wonder.
| Statement | True or false? |
|---|---|
| Generative output approximates patterns in the training data. | ? |
| Real AI thinks the same way a person does. | ? |
| A new-sounding sentence can still be built from old patterns. | ? |
| Generative AI can be wrong. | ? |
Deep thinking. Tomorrow you look for synthetic content in everyday life.