Ask what the model was shown

Computing · Artificial Intelligence · ages 14-15

Name ______________________ Date ____________

A model trained through 2021 is asked about 2024 events. What holds?
- It updates itself nightly
- It knows nothing past its cut-off
- Dates never matter
Which three questions fit a dataset before trusting its model?
- How heavy is it and how blue is the cover
- Which font and which printer
- Who is in it, who is missing, when collected
Whose permission matters when gathering training records?
- The printer
- Nobody at all
- The people the records describe
Blurry photos of some tanks and sharp photos of others taught photo quality, not tanks. What is the lesson?
- Blurry photos always help
- Data quirks become the model's rules
- Tanks cannot be photographed
Why keep a separate test set?
- To train twice as fast
- To delete the training data
- To catch tricks like the tank story before real use
Before trusting a model that was trained on a dataset, which set of questions is most useful to ask about that dataset?
- Who is in the data, who is missing, and when was it collected
- How big the files are, how fast the computer was, and what the model is named
- How many colors are in the charts, who made the charts, and what font was used
- How much the computer cost, where it was stored, and who owns the building
A fitness model trained only on males is used on everyone. What is most likely?
- Equal results for all
- Worse predictions for the missing group
- Better results for the missing group
A group of researchers collected all of their data from males in an athletic association, then used that data to build a model that predicts fitness test results. Later the model is used on everyone, including women and girls. What is the most likely result?
- The model works just as well for everyone, because more math always fixes missing data
- The model tends to make worse predictions for the people who were missing from the training data
- The model refuses to make any prediction for women and girls
- The model works better for women and girls, because it had fewer examples to memorize
Two models read job applications. One trained on all groups with consent, one on a narrow slice. Which earns trust for decisions about people?
- The one trained on all groups with consent
- The narrow one, since small data is cleaner
- Neither needs any data questions
A cut-off date only affects spelling, never facts about events.
Circle one: True False

Answer key

For grown-ups. Fold this page away before handing over the rest.

Ask what the model was shown W1-mt_1Lw0TMHShk-s1

It knows nothing past its cut-off · The model knows nothing past the day collection ended.
Who is in it, who is missing, when collected · Those three answers bound everything the model can do.
The people the records describe · Records about people need those people's permission.
Data quirks become the model's rules · The model learns whatever pattern the data repeats, even the wrong one.
To catch tricks like the tank story before real use · Fresh data exposes wrong lessons before they cause harm.
Who is in the data, who is missing, and when was it collected · What a model can be trusted with depends on who was included in the training data, who was left out, and when the data was collected. Those facts decide whose patterns the model learned and how fresh its knowledge is.
Worse predictions for the missing group · Missing patterns at training mean worse guesses for those people.
The model tends to make worse predictions for the people who were missing from the training data · A model only learns patterns from the examples it was shown. When a whole group is missing from the training data, the model never learns their patterns, so its predictions for that group are usually worse.
The one trained on all groups with consent · Full coverage with permission bounds the model to the people it judges.
False · Events past collection day fall outside everything it was fitted on.