Unplugged machine learning activities for middle school

PD week just ended. A fresh ISTE poster is taped to the hall bulletin board. “Machine learning” is now on the scope-and-sequence — and the district sent exactly zero curriculum to go with it. Twenty-eight students are about to walk in, and the unit plan still has a blank row where “ML lesson” should be. That gap is not a planning failure. Machine learning sits inside the AI4K12 framework and dozens of district AI literacy mandates without a single widely-distributed facilitator guide explaining how to actually teach it. The missing resource layer is the problem. Card sorts and role-plays teach the conceptual layer students need — and none of it requires a single device.

TL;DR: Middle school students can learn machine learning without computers through hands-on card and role-play activities. In supervised learning, students use labeled cards to “train” a sorting rule — like sorting animals by habitat. In unsupervised learning, students group unlabeled cards by finding patterns on their own. These unplugged activities teach the key difference: supervised learning uses labeled examples; unsupervised learning finds structure without labels. No devices needed. Activities run in a 45-60 minute class period and align with AI4K12 Big Idea #3 (Learning) and ISTE 1.5.b and 1.5.c.

What is machine learning?

Machine learning is a type of computer system that improves from examples rather than from explicit instructions. Instead of a programmer writing a rule like “if the email contains ‘click here to win’ then mark it as spam,” a machine learning system reads thousands of past emails that humans already labeled as spam or not-spam — and figures out the pattern itself.

Students already interact with machine learning systems every day. The TikTok recommendation feed learns from which videos they watch all the way through and which they scroll past. Spotify’s Discover Weekly generates a fresh playlist each Monday based on patterns in listening history — no DJ, no explicit playlist rule. The spam filter that stops a phishing email from landing in a school inbox was trained on millions of labeled examples, not a hand-coded checklist. All of those are machine learning systems.

There are two broad categories — supervised and unsupervised learning — and the distinction between them is the conceptual foundation every middle school AI literacy unit needs. This is the heart of AI4K12 Big Idea #3: computers can learn from data. Understanding that foundation is what separates a student who can use AI tools critically from a student who just uses them. For a deeper look at what large language models do with that learned foundation, see the companion post How generative AI works — lesson plan for grade 7.

Supervised vs. unsupervised learning: the key difference your curriculum misses

Every free resource that touches machine learning for middle school describes the mechanics — algorithms, training data, models. Almost none of them name the two paradigms clearly or give students a framework for telling them apart. That’s the gap this section closes.

Supervised learning means the training data comes with labels — an “answer key” the algorithm can check itself against. The algorithm reads example after example, each one pre-tagged by humans, until it can predict the correct label for examples it hasn’t seen before. A spam filter trained on emails labeled “spam” and “not spam” is supervised learning.

Unsupervised learning means the data has no labels. The algorithm has to find its own structure — groupings, clusters, patterns — without any human telling it what the right categories are. Spotify’s genre clustering works this way: no one labeled 10 million songs by hand; the algorithm discovered that certain songs cluster together based on audio features.

	Supervised learning	Unsupervised learning
Training data	Labeled (answer key provided)	Unlabeled (no answers)
Task type	Predict known categories	Discover hidden groups
Student activity	Sort cards using the rule given	Sort cards to find your own rule
Real-world example	Spam filter (spam/not spam labels exist)	Playlist grouping (patterns emerge)
Classroom takeaway	AI learns from human judgment	AI finds structure humans didn’t predefine

That table is the one students need on the desk during the card sort. The Machine Learning Lesson — Supervised vs. Unsupervised Activity (Grades 6-7-8) builds this comparison into a full lesson arc with vocabulary, guided notes, and an exit ticket. Standards anchor: AI4K12 Big Idea #3 — “Computers learn from data.”

How do you teach machine learning without computers?

The answer is a two-round card sort, run back to back in the same class period.

Round 1 (Supervised learning) — the labeled card sort

Each group receives a set of 12 animal-habitat cards. Every card shows an animal on one side; on the back, the habitat is printed — “forest,” “ocean,” “grassland,” or “desert.” That printed label is the training data. Students read through 8 cards, studying the animal features alongside the habitat label, and write down a sorting rule for each habitat in their own words. (“Forest animals tend to have camouflage coloring and are medium-sized.”) Then they flip 4 “test cards” face up — labels hidden — and apply their rule to predict the habitat before checking.

Common student mistake: In Round 1, students memorize individual cards instead of extracting a generalizable rule. They remember “that was the red fox — forest” rather than building the pattern that would work on a card they’ve never seen. Fix: after card 3, prompt groups — “Your rule has to work on cards you haven’t seen yet. What feature are you actually using?” That one sentence redirects almost every group.

Round 2 (Unsupervised learning) — the unlabeled card sort

Same 12 cards. Labels are removed or covered. Groups must now invent their own categories without any guidance — the only instruction is “sort these into groups that make sense to you.” Different groups will produce different groupings: some by size, some by diet, some by body covering. That variation is the lesson.

Debrief question: What sorting rule did your group develop? If you were a machine learning algorithm, what “feature” were you using to make each group? How is discovering that feature different from being told the categories in Round 1?

The debrief is where the conceptual learning lands. Students usually realize, unprompted, that the unlabeled sort produced categories that no one predesigned — and that a real algorithm doing this on millions of data points would find patterns humans might never have thought to look for.

The AI role-play: act out a machine learning system

The role-play takes 45-50 minutes and works especially well with kinesthetic learners or any class that needs to move. Assign four roles around a table or a section of the room:

• Data-input reader — reads raw examples (cards, slips of paper) aloud to the group
• Training algorithm — records patterns, proposes a sorting rule after every 5 examples
• Output predictor — applies the current rule to new examples and announces the predicted label
• Error checker — compares the predicted label to the actual label and reports accuracy back to the algorithm

The algorithm role updates the rule every time the error checker reports a miss. After 20 examples, the group discusses: how did the rule change over time? What would happen with 200 examples? 200,000?

The full 60-card deck and facilitator guide for this role-play is the Unplugged AI Training Simulation (Grades 6-7-8). A no-materials-prep version structured purely around student roles is the No-Tech AI Role-Play — Act Out an AI System (Grades 6-8), which works when printing cards isn’t an option.

Which unplugged activity works for which grade level?

No single activity format works the same way across sixth, seventh, and eighth grade. The table below maps each activity to its best fit — not because the concept changes, but because class time, prior knowledge, and kinesthetic tolerance shift across the middle school band.

Activity	Best for	Class time	Differentiation
Card sort (supervised)	Grade 6 intro	30 min	Add more cards for advanced groups
Card sort (unsupervised)	Grade 7 follow-up	30 min	Limit categories for support groups
AI role-play (P128)	Grade 7-8, kinesthetic learners	45-50 min	Assign simpler data-input roles to support
Paper investigations (P129)	Grade 8 or mixed-ability class	60 min	4 investigations scaffold from concrete to abstract

The paper investigations in that last row are worth calling out separately. A mixed-ability grade 8 class benefits from a scaffolded sequence that starts with a concrete physical sort (Investigation 1), moves to data tables (Investigation 2), introduces error analysis (Investigation 3), and ends with an open-ended design challenge (Investigation 4). That arc is the Low-Tech AI Investigation Pack — 4 Paper Investigations (Grades 6-8). For a broader set of no-tech formats including station rotations, see AI lesson plan — no computers, middle school.

The scope-and-sequence said “machine learning.” It didn’t say which activity fits which grade, how long each one runs, or what the differentiation move looks like for a student reading two years below grade level. That’s the part that was never in the mandate.

What standards do unplugged ML activities address?

Teachers who need to document these activities for admin or an AI literacy audit can map every card sort and role-play to anchor codes — no generic “standards-aligned” language required.

Activity	AI4K12	ISTE	CCSS
Card sort (supervised)	Big Idea #3	1.5.b (collect + analyze data)	—
Supervised card sort	Big Idea #3	1.5.c (build models from parts)	CCSS.ELA-LITERACY.W.7.8 (assess credibility of examples)
AI role-play	Big Idea #3	1.5.c (algorithmic thinking via roles)	—
Paper investigations	Big Idea #3	1.5.b + 1.5.c	CCSS.ELA-LITERACY.W.7.8

AI4K12 Big Idea #3 (Learning) — “Computers learn from data.” This is the foundational standard for any supervised or unsupervised learning activity; every card sort and role-play above lands here.

ISTE 1.5 = Computational Thinker. Two sub-standards apply directly:

• ISTE 1.5.b — students “collect data or identify relevant data sets.” In the card sort, students handle the labeled training set and identify which features carry information — that is data collection in the ML sense.
• ISTE 1.5.c — students “break problems into component parts, extract key information, and develop descriptive models.” The sorting rule students write in Round 1 is a descriptive model. The role-play distributes the model-building across four distinct component roles.

CCSS.ELA-LITERACY.W.7.8 maps to the supervised card sort and paper investigations through the evidence-assessment move: students evaluate which features in the training examples are credible evidence for a sorting rule, which is the same cognitive operation as assessing source credibility in a research task.

Standards source: ISTE Student Standards. ISTE is a registered trademark of the International Society for Technology in Education. These resources are not affiliated with or endorsed by ISTE.

Ready to print before Monday

The scope-and-sequence said “machine learning.” It didn’t come with a facilitator guide or a set of 60 sorting cards. That gap was never yours to fill from scratch.

Three starting points depending on where your class is this week:

• Machine Learning Lesson — Supervised vs. Unsupervised Activity — the concept-first lesson that teaches the two paradigms with guided notes and the comparison table. Best first move for a class that hasn’t heard either term.
• Unplugged AI Training Simulation — 60 Cards — the full card deck with facilitator guide. Print, cut, hand out. A typical class finishes the two-round sort plus debrief in 55 minutes.
• Low-Tech AI Investigation Pack — four scaffolded paper investigations, ready to assign Monday, no prep beyond printing.

Browse the full no-tech AI literacy catalog at /shop or start with free materials at /free.

This post was drafted with AI assistance and human-finalized.