← Back to course
AI and You 9-12 / Week 04 / Tuesday
2/6
Week 04 · How Generative Models Write

Tuesday

Smooth but wrong
// One word at a time, smooth but not always true
⏱ about 20 min

Tuesday: Smooth but Wrong

Comet fills a whole page of her sketch notebook with Tiny Text sentences. "I think I have found every one it can make," she says.

Wren counts them. "Thirteen. But only six training cards, and two of those are the same."

"So where did the other sentences come from?" Comet asks.

Nova hovers over the page and projects three boxes: in the data, contradicts the data, unsupported. "Would you like a hint?" she asks. "Check each sentence against the cards, one at a time."

Comet reads "the lab has quiet rooms" aloud. "That sounds totally fine."

Wren flips through the cards. "No card says anything about quiet rooms in the lab."

Comet frowns. "So it sounds right, but nothing backs it up."

Every sentence Tiny Text can write

Tiny Text can write 13 different sentences. Nine end with "opens at" and an hour. Four use "has".

The crew sorts each one into three boxes.

BoxWhat it means
In the dataA training card says exactly this.
Contradicts the dataA training card says something different.
UnsupportedNo card says anything about it either way.
NumberPossible sentenceBox
1the library opens at eightin the data (T1, T5)
2the library opens at ninecontradicts (T1 says eight)
3the library opens at sevencontradicts (T1 says eight)
4the lab opens at eightcontradicts (T3 says nine)
5the lab opens at ninein the data (T3)
6the lab opens at sevencontradicts (T3 says nine)
7the gym opens at eightcontradicts (T6 says seven)
8the gym opens at ninecontradicts (T6 says seven)
9the gym opens at sevenin the data (T6)
10the library has quiet roomsin the data (T2)
11the library has new desksunsupported
12the lab has quiet roomsunsupported
13the lab has new desksin the data (T4)

This is a made-up example: the crew's own sort of Tiny Text.

COUNT THE BOXES
  • Read the question.
  • Tap your answer.
How many of the 13 sentences are in the data?
How many contradict the data?
How many are unsupported?
So how many of the 13 sentences are new, not on any card?

Confabulation

Confabulation is when generative AI confidently presents wrong or false content.

It also covers outputs that wander from the question, or that contradict what the system said earlier.

Confabulation is a natural result of how generative models are designed.

The same kind of prediction can give a true output or a false one. Nothing in the sentence marks which is which.

Think about it
Tiny Text never says "maybe." Every sentence comes out in the same plain, sure voice.
It joined pieces of true sentences into new ones. Some of the new ones happen to be false.
SORT THESE SENTENCES
  • Read the question.
  • Tap your answer.
"the gym opens at nine." Which box?
"the library has new desks." Which box?
"the lab has new desks." Which box?
StatementTrue or false?
Confabulation means confidently presenting wrong or false content.?
Confabulation only happens when a model is broken.?
The same prediction can give true or false outputs.?
A smooth, confident sentence is always a true sentence.?
WHY THIS EXERCISEKnowing confabulation comes from the design helps you check every output, not only odd-looking ones.
What is the word for when generative AI confidently presents false content? Type it.
WHY THIS EXERCISENaming the risk lets the crew write it into SH-1's risk register as R1.

Sharp sorting, reviewer. Tomorrow in the Explorer Lab you will roll ten sentences of your own.

← Monday