This original 35-minute paper lesson separates a target policy, labeled examples, a hand-written candidate rule, and new test cases. It uses an analogy to discuss classification and evidence. No AI system is trained or evaluated, and actual learner explanations remain blank until teaching occurs.
Code.org's US-facing AI curriculum describes hands-on work with predictions, categories, and sorting, alongside separate activities involving real machine learning. That supports the teaching context; the cards and rules below are independent editorial examples, not its game or a claimed curriculum outcome. Code.org's AI curriculum overview
State the target and show the limited examples
The classroom target is explicit: keep squares and omit circles, regardless of color. Prepare original paper cards or complete written card descriptions. Keep the target statement separate from the labels and any proposed rule.
| Example ID | Card description | Supplied target label |
|---|---|---|
| T01 | Blue square | Keep |
| T02 | Blue square | Keep |
| T03 | Red circle | Omit |
| T04 | Red circle | Omit |
Both color and shape match these four labels. They alone do not establish that color is the appropriate property for the target. Duplicating a blue square and a red circle has not added the contrasting combinations that would distinguish the two hypotheses.
The original candidate C01 is hand-written: blue → Keep, red → Omit, other color → Unknown. Label it as a candidate to test, not as a trained model's prediction. Ask learners what it will do before they compare it with the target key.
Test with separately identified cards
Use four new illustration cards that were not among T01–T04:
| Test ID | Card | Target | C01 prediction | Comparison |
|---|---|---|---|---|
| E01 | Red square | Keep | Omit | Incorrect |
| E02 | Blue circle | Omit | Keep | Incorrect |
| E03 | Green square | Keep | Unknown | No decision |
| E04 | Green circle | Omit | Unknown | No decision |
C01 produces zero correct matches, two incorrect predictions, and two Unknown results on these four cases. Unknown is not a correct prediction. Preserve all three counts and the individual rows; do not present two known-color predictions as two successes.
The counts describe this hand-authored toy rule and these supplied cards. They are not real ML accuracy, measured classroom performance, or evidence about all classification systems. Keep a clean learner test log separate from the teacher table, with blank fields for prediction, target comparison, and explanation.
Add relevant contrast without rewriting the first test
The original clarification adds T05, a red square labeled Keep, and T06, a blue circle labeled Omit. Keep T01–T04 unchanged. These contrasting examples directly expose why color alone cannot satisfy the stated task.
Candidate C02 uses the supplied target: square → Keep, circle → Omit. Its expected answers match all four E cards. This is a teacher key for a new hand-written rule. Comparing C02 on those same known test cards is a revision demonstration, not an independent generalization test. Further relevant evidence would need new cases appropriate to the actual task.
Ask learners why E01 and E02 refute C01 and what the green-card Unknown state means. Do not teach that adding a few examples automatically fixes every real AI error or bias. The activity illustrates a reasoning distinction; actual machine-learning behavior involves a system that this paper task has not created.
Use the 35 minutes for predictions and explanations
Spend five minutes on the objective and target policy. Learners identify the property named by the task. Give eight minutes to inspect T01–T04, propose apparent patterns, and record C01 explicitly.
Use ten minutes for the separate E-card test and error log. Learners apply the candidate before comparing the supplied target labels, retaining incorrect and Unknown results separately. During seven minutes of revision work, introduce T05/T06 and contrast C02 without erasing the first-test record.
Use the final five minutes for an exit explanation: “Which two tests refute the color-only rule? Why does adding relevant data not guarantee that every real AI result is correct?” The phases total 35 minutes. Review the actual explanation rather than automatically grading prose for a keyword. Teacher answers and hypothetical discussion points are not learner responses.
The reusable packet consists of original example/test cards, target and label key, candidate-rule comparison, blank first-test log, distinct revision record, and limitations explanation. It needs no model account, external API, student data, or mandatory computer activity.
Copy this prompt for the classification lesson packet
Create a complete US English 35-minute lesson teaching AI-related classification reasoning through an explicitly hand-authored paper analogy. Return teacher plan, original printable card descriptions, target/label key, candidate comparison, clean learner first-test log, distinct revision record, and exit explanation prompts. No actual ML training, model/API/accounts, student data or copied provider game/artwork.
Target: keep squares, omit circles regardless of color. Initial examples T01 blue square Keep; T02 blue square Keep; T03 red circle Omit; T04 red circle Omit. Both color and shape fit these four labels; they do not justify color as the target property.
Hand-written C01: blue Keep, red Omit, other color Unknown. Separate test cards: E01 red square targetKeep/C01Omit incorrect; E02 blue circle targetOmit/C01Keep incorrect; E03 green square targetKeep/C01Unknown; E04 green circle targetOmit/C01Unknown. Counts zero correct, two incorrect, two unknown. Unknown notcorrect. Preserve first ledger, not a real model accuracy result.
Add original contrasting examples T05 red square Keep and T06 blue circle Omit without changing initial records. Target-based hand-written C02 squareKeep/circleOmit gives four expected target matches. Retesting same E cards is not independent generalization; no automatic bias cure or universal AI-correctness claim.
Timing:5 objective/target,8 examples/hypotheses,10 separate test/log,7 additional examples/revised rule,5 exit=35. Exit asks which E01/E02 cases refute color and why relevant data does not guarantee all real AI results. Actual learner predictions/explanations blank, teacher key separate; no automatic open-response grading.
Proposed local HTML first/revised/key views only display fixed rules/records and collect nothing; construction/testing pending, no training or fake universal AI score. Full text/color-and-shape labels required. Actual learner work stays in the teaching workflow.
Review the examples and test log in NOS
NOS can save and render a prepared HTML copy of the fixed cards, candidate rules, and separate test records. Use the local HTML example to compare the first incorrect and Unknown results with the separate revision record. A view switch displays supplied logic; it does not train AI or establish learner outcomes.
Finish with learners explaining the two counterexamples and the limits of the revision demonstration, so a rule that fits familiar cards is not mistaken for evidence that every new case will be correct.