Skip to content
Idaho AI Classroom
All lessons
AI FluencyEvaluation and Critical Thinking9-12.AIF.ECT.2

Patterns In, Patterns Out

9–1050 minutes90 minutes (extend hands-on and discussion)

Standard quoted exactly

Recognize and understand that AI systems are a technology that uses patterns in data to make decisions or generate new things.

Example from the standards. LLMs were trained on older writings that used the em dash consistently, hence why LLMs rely on the em dash so frequently.

Student-friendly learning targets

  • I can explain that AI systems learn patterns from data and then use those patterns to decide or to generate new text, images, or labels.
  • I can show how a small training set changes the next word a paper 'model' predicts.
  • I can connect the em-dash example to other copied habits in AI output.

Essential questions

  1. If AI is not magic, what is it actually doing with data?
  2. Why does a model repeat habits it saw in training, including punctuation and cliches?
  3. What happens when the patterns in the data do not match Idaho life?

Objectives

  1. Students will describe AI as pattern-matching technology rather than as a mind or a search engine.
  2. Students will complete a next-token (next-word) task by hand and compare it to a printed LLM completion.
  3. Students will explain the standards example about em dashes as a copied training habit.
  4. Students will give two Idaho examples of pattern-in, pattern-out systems (agriculture, manufacturing, health, or recreation).
  5. Students will name a limitation that follows from training only on past data.

Key vocabulary

Pattern
A regularity in data, such as which words tend to follow which, or which pixel groups tend to mean 'bruise on a potato.'
Training data
The examples a model is shown so it can learn those patterns. The mix of examples shapes later output.
Machine learning
A way of building AI in which the system improves on a task by finding patterns in large datasets instead of following a hand-written rule for every case.
Large Language Model (LLM)
A generative model trained on vast text so it can predict likely next tokens and produce human-like writing.
Generative AI
AI designed to create new content (text, images, audio, code) by sampling from learned patterns, not by retrieving one stored document.
Token
A chunk of text, often a word or part of a word, that a language model predicts one step at a time.
Prediction
The model's guess about the most likely next token, label, or value given the patterns it learned.

Teacher background

Students treat AI as magic or as a giant lookup table until they see the simpler claim in the standard: these systems use patterns in data to make decisions or generate new things. An LLM does not know Idaho. It has seen huge numbers of word sequences and learned which tokens tend to follow which. That is why models lean on the em dash. Older professional writing in the training mix used it constantly, so the model copies the habit. The same logic holds off the page. A potato packing line that flags bruises from photos is matching visual patterns. A Micron defect classifier does the same with wafer images. A hospital staffing tool forecasts weekends from past Fridays. None of them understands a tuber, a chip, or a nurse. Ninth- and tenth-graders can hold this without calculus. Use paper next-word tasks, a short projected completion, and a printed packet so the class still runs when a site is blocked. End by naming the human who still chooses the data, the prompt, and whether the output is good enough to use on a farm, a line, or a floor.

Materials and prep

Materials

  • Projector for a teacher-typed completion demo (district-approved tool) or slides of a saved completion
  • Printed fake-output packet: Idaho paragraph with blanks, em-dash example, and a 20-sentence mini training set
  • Word cards or cut strips for a tabletop next-word model
  • Chart paper labeled Patterns in / Patterns out
  • Chromebooks optional; paper packets required for every student
  • Timer, whiteboard, exit tickets

Before class

  • Print the fake-output packet. Highlight the standards em-dash example on page 1.
  • Build a 20-sentence mini set that overuses one habit (for example, starting sentences with 'Overall,' or inserting em dashes).
  • If live demo is allowed, test the class prompt on the teacher account only. Save a screenshot as the offline fallback.
  • Queue two images: bruised vs. clean potatoes, or a ski-report chart, to make non-text patterns visible.
  • Decide pairings in advance; the hands-on needs talk, not silent guessing.

Instructional sequence

Finish the sentence

5 min
  1. Project: 'After a dry July, the Sawtooth forest is more likely to ___.' Students write one word on the packet.
  2. Tally the top three class guesses. Note that we are already doing pattern completion from Idaho news we have heard.
  3. Reveal a printed LLM completion that uses an em dash and a generic phrase. Ask what habit was copied.

Patterns in, patterns out

10 min
  1. Write the standard in plain language: data in, pattern learned, decision or new thing out.
  2. Teach the em-dash example exactly as written in the standards. The model is not fond of punctuation; it is replaying frequency.
  3. Show a non-text twin: potato bruise photos in, 'divert this tuber' out; Micron images in, 'hold this wafer' out.
  4. Name what the system does not do: it does not visit the field, smell smoke, or understand a sentence the way a student does.

Next word, by hand

8 min
  1. Read the packet's five-sentence Magic Valley irrigation paragraph. Cover the last word of each sentence.
  2. Pairs predict the next word using only patterns in those five sentences, not outside knowledge.
  3. Uncover. Compare pair guesses to the printed LLM completion. Mark copies of training habits (repeated openers, em dashes, 'delve').

Explain the em dash

10 min
  1. Students write a six-sentence explanation of the standards example for a seventh grader: what was in the training writing, what the model does at generation time, and why frequency is not understanding.
  2. Add one Idaho parallel: a snow-report model trained only on Colorado resorts, or a yield model trained only on Iowa corn.
  3. Collect a mid-lesson check: star the sentence that would be false if AI 'looked up' a single stored article instead of generating from patterns.

Real-world examples

  • Potato packing plants in the Magic Valley can train vision models on bruise photos; the label 'bruise' is a pattern, not a farmer's judgment in the cellar.
  • Micron in Boise uses pattern recognition on wafer images to flag defects that look like past defects.
  • Hospital census tools predict busy nights from historical Fridays and flu weeks, not from knowing a particular patient.
  • A recreation snow-report bot trained on national ski writing may sound right while missing Bogus Basin or Brundage habits.
  • County IT spam filters learn patterns in phishing mail; a new local scam that does not match the pattern can slip through.

Hands-on activity

Build a 20-sentence 'model'

12 min
  1. Each pair gets 20 printed sentences that overuse one habit (em dashes, 'Overall,', or 'in today's fast-paced world').
  2. Students tally the habit, then generate a five-word continuation of a new prompt using only words and habits from the set. Write it on the packet.
  3. Compare the tabletop output to the printed LLM sample on the last page. Circle shared habits.
  4. Teacher projects a live completion only if the district tool is up; otherwise project the saved screenshot. The paper model is the required product.
  5. Close the activity by answering: if we trained on Idaho extension bulletins instead of old national magazines, what would change?

Discussion questions

  1. Why is 'the model likes em dashes' a misleading sentence?
  2. What would a wildfire model copy if it were trained only on California fire seasons?
  3. How is generating a new sentence different from searching a library catalog?
  4. Where should a trades apprentice still refuse to trust a generated procedure even if the sentences look familiar?
  5. What pattern in school essays might a homework detector be copying, for better or worse?

Differentiation

Support

  • Provide a cloze frame: 'The model saw ___ often, so it ___.'
  • Let students work with a 10-sentence set instead of 20.
  • Offer a spoken explanation option recorded to the teacher, not to an AI tool.

Challenge

  • Ask students to design a tiny training set that would make a model overuse Idaho place names, then warn about the downside.
  • Compare predictive AI (next staffing number) to generative AI (next paragraph) using the same pattern logic.
  • Find a second copied habit in the printed LLM sample besides the em dash.

Multilingual learners

  • Point out that a model trained mostly on English will copy English habits; home-language patterns may be weak or stereotyped.
  • Allow the seven-grader explanation to be drafted in the stronger language, then translated to English with a partner.
  • Pre-teach token, pattern, and generate with gestures: stack of examples, then a new card.

IEP / 504

  • Enlarge the sentence cards and limit the tally to one target habit.
  • Accept a labeled diagram (data in, pattern, output out) in place of six sentences.
  • Provide the independent write as bullet prompts rather than a paragraph.

Assessment

Formative

  • Warm-up word and the next-word pair task.
  • Independent paragraph on the em-dash example.
  • Hands-on continuation that clearly borrows a training habit.

Summative

  • Exit ticket: Complete 'AI uses ___ in data to ___ or ___.' Then give one Idaho example and one limitation of training on the past.
  • Packet page scored for a correct contrast between pattern generation and looking up a stored document.

Success criteria

  • I described AI as a pattern technology, not a mind and not a simple search box.
  • I explained the em-dash example as a copied habit from training data.
  • I showed how changing the tiny training set changed the output.
  • I named a human who still chooses data and checks results.

Responsible use, ethics, and privacy

Responsible use

A class prompt on a teacher account is enough. Students should not paste their own essays into an unapproved generator to 'see the pattern.' The paper model teaches the idea without creating a cheating shortcut.

Ethics

Copied patterns can spread stereotypes as easily as punctuation. If the training writing assumed doctors are men or Idaho is only Boise, those habits come out too. Later lessons on bias start from this mechanism.

Privacy

FERPA: no student PII in unapproved tools. Do not upload class essays, names, or school photos to a public model to demonstrate patterns. Use the printed mini-set and teacher-made samples. Chromebooks, if used, stay on the district tool with a shared prompt.

Reflection

  1. What habit would my own text messages teach a model?
  2. Where have I seen an AI output that felt familiar but empty?
  3. Which Idaho dataset would I want in the training mix before I trusted a local prediction?

Homework

On paper, copy three sentences from a public source (textbook, newspaper, or today's packet). Circle a repeated habit. Write five sentences explaining what an LLM might copy from that source, using the em-dash example as a model. No new AI accounts.

Closing

Read the standards example aloud. Students complete the oral stem: 'Patterns in, ___ out.' Collect packets. Tomorrow or next ECT lesson will ask what happens when those patterns are biased or simply wrong.

Extensions and cross-curricular links

Go further

  • 90-minute block: run a second mini-set (image labels for potato bruises or weld quality) and compare text patterns to visual patterns.
  • Invite an ag teacher or extension educator to show how yield models depend on local data, not national averages.
  • Math connection: relative frequency of a punctuation mark in a 20-sentence set; no calculus required.
  • Students collect three public AI-looking sentences and highlight copied filler habits, on paper only.
English Language Arts
Style and voice: overused dashes, 'delve,' and stock openers are patterns, which is why they show up in generated essays.
Agriculture / Plant Science
Vision models on packing lines only recognize defects they have seen in training photos.
Statistics
Frequency in a sample is not understanding; it is a count that can mislead when the sample is the wrong world.
Computer Science
Next-token prediction as a gentle preview of how LLMs generate, without requiring code.