Children can learn how machines learn through hands-on activities that start with sorting examples and end with testing a trained model. The AI4K12 initiative's draft progression chart for its Learning topic, version 0.1 released November 19, 2020, sets concrete goals for each grade band, such as learning a made-up class called "poisonous fish" from labeled pictures in kindergarten through grade 2. Evidence that such activities work is thinner than the activities themselves: a 2023 review of 179 documents found hardly any that tested whether students understood the concepts afterward.
Key takeaways
- The AI4K12 chart is a working draft that organizes goals into four grade bands: K-2, 3-5, 6-8, and 9-12.
- Early grades focus on examples and rules (sorting, guessing, labeling), and later grades add decision trees, accuracy tests, and tabular datasets.
- A systematic review of 179 documents reported hardly any studies that assessed student understanding after a learning experience.
- Good activities make children check a model on new examples, since that is where the limits of machine learning show up.
What the draft chart is and what it is not
AI4K12 publishes progression charts that list, for each concept, a learning objective (what students should be able to do) and an enduring understanding (what they should know). The chart for the Learning topic opens with its central claim: computers can learn from data, and machine learning lets a computer acquire behaviors without people explicitly programming them. The draft chart is labeled subject to change after public feedback, so it is a set of suggestions from a working group, not a curriculum rule in any province or state.
Its value for adults planning an activity is specificity. Instead of saying children should "understand AI," it names tasks that can be done with paper, a webcam, or a free classroom tool.
Kindergarten through grade 2: examples, patterns, and one simple model
The chart asks young children to describe how people learn and how computers learn, and to see that computers learn differently. Its suggested discussion is to recall a time they learned something by being told, by watching someone, or by asking questions. A demo with positive and negative examples can then show a computer learning from them.
For pattern finding, the chart proposes a class that is not obvious on purpose: children look at cartoon fish labeled "poisonous" or "not poisonous" and say which features seem to predict the label, such as a red body with a square head. Because the rule is arbitrary, children cannot lean on common sense and have to find the pattern in the examples. The chart also lists training a computer to recognize something, such as hand gestures or sounds, with an instructor's help.
Grades 3 to 5: labeling, decision trees, and checking accuracy
The next band moves from saying a rule to drawing it. Students sketch a decision tree for the fish task and then face harder classes, such as fish that count as poisonous if they are red with square heads, blue with round heads, or purple with pointy spines. The chart also asks students to train a classification model, then examine its accuracy on new inputs and note that more training may be needed if results are poor.
It adds a labeling exercise: define the classes first, then add examples for each class, and observe that a model shown unlabeled new data predicts a class without getting any feedback on whether it was right. Game-style options in the chart include a "guess the animal" game that builds a decision tree and a grid-world demonstration of trial-and-error learning, where an agent finds the best move in each square over repeated tries.
Grades 6 to 12: data tables, test sets, and evaluation
For grades 6 to 8, the stated goal is to train and evaluate a model on a tabular dataset, where each row is an example and each column a feature, and the learning algorithm is likely a decision tree. Grades 9 to 12 are asked to train a supervised or unsupervised model on real-world data and measure the percent of test items labeled correctly. The chart notes that younger students might test a model on a few new items, while this band is expected to measure performance on a nontrivial test set.
Activity map from the draft chart
| Grade band | Activity named in the chart | Skill practiced |
|---|---|---|
| K-2 | Labeled cartoon fish ("poisonous" or not) | Finding features that predict a label |
| K-2 | Teacher-assisted recognition of gestures or sounds | Computers learning from examples |
| 3-5 | Drawing a decision tree for harder fish classes | Writing rules as a tree |
| 3-5 | Training a classifier and testing it on new input | Accuracy and retraining |
| 3-5 | "Guess the animal" tree-building game | How a learner adds nodes after a wrong answer |
| 6-8 | Training and evaluating on a table of features | Rows, columns, and classification |
| 9-12 | Supervised or unsupervised model on real data | Percent correct on a test set |
What the evidence says about benefits
Claims that early AI lessons build skills usually rest on enthusiasm, not tests. The 2023 systematic review in the International Journal of STEM Education, by Casal-Otero and colleagues, searched Scopus and reviewed 179 documents on AI literacy in K-12. It found two broad groups, learning experiences and theoretical frameworks. It also reported that there were hardly any experiences that assessed whether students understood AI concepts after the lesson, and that little attention had gone to undesirable consequences of applying AI without careful thought.
That does not show the activities fail. It shows the field has mostly described what was done, not measured what children kept. A school or family that wants evidence should build in its own small check, such as asking a child to predict how a model will handle a new example and then testing the prediction.
Checklist for a home or classroom activity
- Does the child supply or label the examples, not only watch a finished demo?
- Is the model tested on examples it has not seen?
- Does the activity show a mistake, and does the child explain why it happened?
- Does it avoid sharing names, photos, or voices of children with outside services without parental consent and school approval?
- Is the chosen activity matched to the grade band, with the age guidance treated as a suggestion?
FAQ
At what age can children start learning how AI works?
The draft chart begins in kindergarten through grade 2, with discussion and example-sorting tasks. Its goals are stated as concepts, and none of the K-2 objectives mentions writing code.
Does a child need to code to learn about machine learning?
The chart's early bands describe discussion, spoken rules, drawn decision trees, and teacher-assisted demos, and none of its objectives for those bands mentions writing code. Tabular datasets and test sets enter in grades 6 to 8.
Is the AI4K12 chart official policy?
No. It is a draft from a working group and is marked subject to change. Ministries, provinces, and states set their own requirements.
Are there proven learning gains from teaching AI to kids?
The 2023 review found hardly any studies that checked student understanding, so gains are largely untested. Short, checked activities are a reasonable way to find out locally.
Related reading
- What to look for in AI camps for kids and teens
- How interactive AI systems work
- AI activities for language learners
- What AI toys do and what to check
More in AI in education.
Last verified: October 2026
Sources
- Draft Big Idea 3 Progression Chart, version 0.1, AI4K12, November 19, 2020
- AI literacy in K-12: a systematic literature review, Casal-Otero et al., International Journal of STEM Education, 2023