Librarian Joss tapes a new sign to the library door: "Now opening at nine." The crew stops in the hallway to read it.
Comet pulls out the Tiny Text tally table. "Uh oh. Our model still says the library opens at eight."
Wren checks the tallies. "What does the evidence say? Nothing in our cards has changed. The model only knows its training data."
"So we just add two new cards," Comet says, already writing.
Nova projects the old cards and the new ones side by side. "Would you like a hint?" she asks. "Count again. Then ask what happens to the old cards."
Wren turns to you. "Reviewer, how do we fix Tiny Text the right way?"
Training data can go stale. It can fall out of date for the place where a system is used.
AI systems may be trained on data that changes over time. That can change how well they work in ways that are hard to understand.
So AI systems may need fixing more often than ordinary software.
Tiny Text was trained when the library opened at eight. The world changed, but its training data did not.
Comet's first fix adds two new cards, both saying "the library opens at nine," and keeps the old ones.
Wren's fix swaps T1 and T5, the two old library-hours cards, for the two new cards.
| Version | After "at" (tally) | Row total | Most likely after "at" |
|---|---|---|---|
| Original Tiny Text | eight 2, nine 1, seven 1 | 4 | eight |
| Comet's fix: add two new cards, keep T1 and T5 | eight 2, nine 3, seven 1 | 6 | nine |
| Wren's fix: swap T1 and T5 for the two new cards | nine 3, seven 1 | 4 | nine |
AI systems depend on their training data, which is often huge and complex.
Changes made during training, on purpose or by accident, can change how a system performs.
Swapping two cards changed what Tiny Text says. A real system has far more data, so a change can be much harder to trace.
Tiny Text also looks only one word back. So after "at," it cannot tell the library from the gym.
| Statement | True or false? |
|---|---|
| Training data can go stale. | ? |
| Once a model is trained, the world can never make it wrong. | ? |
| Changes during training can change how a system performs. | ? |
| Adding new cards always removes the old answer. | ? |
| AI systems may need fixing more often than ordinary software. | ? |
Careful fixing, reviewer. Tomorrow is Impact Friday: what can we know about SH-1's own training data?