Skip to content
Idaho AI Classroom
All lessons
AI FluencyAppropriate Use9-12.AIF.AU.4

Not All Models Speak

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

Standard quoted exactly

Understand that Large Language Models (LLM) are just one type of AI model and that others exist with different purposes.

Example from the standards. LLMs are trained on natural language text and generate text, whereas diffusion models are used for image generation, decision tree models can classify, linear regression models can predict numeric values, etc.

Student-friendly learning targets

  • I can explain that an LLM is only one kind of AI model and that other kinds have other jobs.
  • I can match a sample output to a model family: language, image, classification, or numeric prediction.
  • I can name a task each family should not be asked to do.

Essential questions

  1. If a chatbot is only one kind of AI, what are the others for?
  2. How should the type of output (text, image, category, number) guide which model we trust?
  3. Why is “just ask the AI” a weak plan when models are built for different purposes?

Objectives

  1. Define LLM, diffusion model, decision tree, and linear regression at a grade 9–10 level.
  2. Match mystery outputs to model families using purpose, not brand names.
  3. Give an Idaho example for at least two families (yield numbers, wafer classification, trail image, permit letter).
  4. State one misuse for each family (for example, using an LLM as a calculator of record).
  5. Read a tiny linear pattern and a two-split tree on paper so the math is visible, not magical.

Key vocabulary

Large Language Model (LLM)
A model trained on huge amounts of text so it can predict and generate language. It is one type of AI, not all of AI.
Diffusion model
A model used to generate images (and sometimes audio or video) from patterns it learned, often from a text prompt.
Decision tree
A model that classifies by asking a sequence of yes/no or threshold questions, like a flowchart.
Linear regression
A model that fits a line (or a simple trend) to data in order to predict a number, such as yield or temperature.
Classification
Assigning an item to a category (in spec / out of spec, frost risk / no frost) rather than writing a paragraph.
Prediction
A model’s estimate of a future or unknown value, based on patterns in past data—not a guarantee.
Training data
The examples a model is shown during building. Different families need different kinds of examples (text, images, labeled rows, numbers).
Model family
A group of models that share a purpose and a kind of output, the way trucks, boats, and planes are all vehicles with different jobs.

Teacher background

Chatbots dominate the news, so students assume AI means “the thing that talks.” The standard requires a wider map: language models generate text; diffusion models generate images; decision trees classify; linear regression predicts a number. Ninth graders can hold those four families if each one is tied to a job they recognize. Use Idaho numbers, not abstract ones. Linear regression can estimate potato yield from soil moisture. A decision tree can flag whether a Micron wafer lot is in spec. A diffusion model can invent a “photo” of a Sawtooth trail that does not exist. An LLM can draft a county permit letter that sounds official and is wrong. Pitfalls: drowning the period in architecture diagrams, or implying one family is “smarter.” Purpose, not prestige, is the sorting rule. Keep math honest and light—show a three-point line and a simple tree on the board. Project a district-approved demo only after students have matched printed “mystery outputs” to model types on paper. That packet is the lesson when the network dies. Students should leave able to name the model family, the kind of output, and a task it should not be asked to do.

Materials and prep

Materials

  • Four family posters: LLM (text), diffusion (image), decision tree (category), linear regression (number)
  • Mystery-output packet: six printed outputs with no labels (letter, fake trail photo, yield table, wafer pass/fail tree, chatbot poem, frost-risk flowchart)
  • Board-ready three-point data set (soil moisture vs. yield) and a two-split tree (temp / moisture → frost risk)
  • Matching half-sheet (one per student)
  • Optional teacher demo on a district-approved tool: one text prompt and one image prompt, teacher account only
  • OFFLINE fallback: the mystery-output packet and board math replace all live tools

Before class

  • Print mystery outputs. Caption the generated image as “generated, not a photograph.”
  • Practice the three-point line so you can plot it in under a minute.
  • If demoing, script two prompts that use public facts only (no student work).
  • Decide a local number students know (cwt per acre, river cfs, or a shop measurement) to stand in if potato yield feels distant.

Instructional sequence

Four outputs, no brands

7 min
  1. 1. Reveal four unlabeled products: a paragraph, a picture, a pass/fail tag, and a predicted number.
  2. 2. Students silently write which one “looks like AI” to them and why. Most will pick the paragraph or the picture.
  3. 3. Tell them all four can be AI, and that today’s job is purpose, not brand logos.
  4. 4. Introduce the sentence we will reuse: “This model is for ____, so I would not ask it to ____.”

Four families, four jobs

13 min
  1. 1. Write the standard and the standards example on the board. Read them exactly.
  2. 2. Map four families: LLM → text; diffusion → images; decision tree → classification; linear regression → numeric prediction.
  3. 3. Plot three points (moisture vs. yield) and draw a rough line. Predict a fourth yield. Say the word “estimate.”
  4. 4. Draw a two-split tree: temperature below 32? moisture high? → frost-risk class. Say the word “classify.”
  5. 5. Misuse examples: an LLM is a poor calculator of record; a diffusion model is a poor witness; a regression line is a poor poem; a tree is a poor substitute for a full diagnosis at St. Luke’s.

Match with a purpose sentence

10 min
  1. 1. Pairs open the mystery-output packet and place each of the first four items on a family poster.
  2. 2. For each, they write: “This model is for ____, so I would not ask it to ____.”
  3. 3. Teacher circulates for two common errors: calling every paragraph an LLM (could be a person) and calling every picture a photograph.
  4. 4. Confirm as a class. Leave two mystery items unsolved for independent work.

Two remaining mysteries plus a misuse

10 min
  1. 1. Each student matches the last two packet items and writes the purpose sentence for both.
  2. 2. They choose one Idaho workplace (potato yield, Micron lot, county letter, trail tourism image) and name the best family and a family they would refuse.
  3. 3. Exit box: “An LLM is only one type of AI because ____.”
  4. 4. No student accounts. If a neighbor is stuck, they may borrow a family poster, not a chatbot.

Real-world examples

  • Linear regression on soil moisture and historical yield can estimate cwt per acre for a Magic Valley grower. It will not write the crop-insurance narrative.
  • A decision tree can classify Micron wafer lots as in spec or out of spec. It should not generate a press photo of the clean room.
  • An LLM might draft a county permit letter that sounds official. A diffusion model might invent a Sawtooth “photograph” for a tourism page. Neither is a substitute for a surveyor or a camera on the trail.

Hands-on activity

Mystery desk and optional demo

12 min
  1. 1. Paper core: teams of three play “wrong tool.” The teacher names a job (predict tonight’s low in Arco; generate a poster; classify smoke plumes as hold or go; write a parent email). Teams hold up the family card they would refuse and say why.
  2. 2. Teams then hold up the family they would consider, with a human check named.
  3. 3. Optional digital: teacher runs one text prompt and one image prompt on a district-approved tool. Class labels the family before the result appears.
  4. 4. If the network is down, skip step 3. Use two extra printed outputs in the packet.
  5. 5. Close by reciting the standards example in chunks; students supply the family names in unison.

Discussion questions

  1. Why do news stories talk as if “AI” were one product?
  2. Which family is easiest to over-trust because the output looks human?
  3. When would a simple tree or a line on paper beat a huge language model?
  4. What should a county website or a hospital dashboard label so the public knows which family produced the result?

Differentiation

Support

  • Give a four-row cheat sheet: family, output type, Idaho example, common misuse.
  • Allow matching with sticky notes onto posters instead of written purpose sentences; then scribe one sentence together.
  • Color-code packets (blue text, green image, yellow category, orange number).

Challenge

  • Add a fifth family in a short research note (clustering, recommendation, or object detection) using a printed excerpt.
  • Explain why an LLM can emit a number that is not a regression prediction.
  • Design a one-page “model menu” a potato shed or a county clerk could actually post.

Multilingual learners

  • Keep family names in English; allow the purpose clause in the student’s strongest language.
  • Use icons on every poster (speech bubble, picture frame, flowchart, line graph).
  • Pre-teach output, classify, predict, and generate with gestures and examples.

IEP / 504

  • Reduce independent mysteries from two to one plus the exit box.
  • Provide a large-print packet and allow oral matching with a teacher or paraeducator.
  • Skip live demos if screen motion is a trigger; paper outputs are the full lesson.

Assessment

Formative

  • Warm-up “what looks like AI” notes and the first four matches during guided practice.
  • Wrong-tool card holds during hands-on (refusal plus a reason).
  • Exit box: “An LLM is only one type of AI because ____.”

Summative

  • Matching half-sheet plus one Idaho workplace item naming a best family, a refused family, and a human check.

Success criteria

  • I named at least three model families and the kind of output each is for.
  • I matched mystery outputs using purpose, not company names.
  • I stated a misuse for at least one family.

Responsible use, ethics, and privacy

Responsible use

Students do not need and should not open image generators or public chatbots for this lesson. Mystery outputs are printed and labeled when synthetic. A teacher demo, if used, runs on a district-approved account with public prompts only. The responsible habit is to ask “what family is this, and what is it for?” before pasting a question into a box. Using an LLM as a silent calculator, or a generated image as a record of a place, is a purpose error, not a clever shortcut.

Ethics

Different families create different harms. A wrong yield number can idle a crew. A wrong classification can scrap a good wafer or release a bad one. A generated image can move tourism money or spread a fake disaster. Fluent text can sound like a county order. Ethics here is matching power to purpose and refusing to hide the family behind the word “AI.” Students also note that training data differ: text scraped from the web is not the same as a grower’s moisture logs, and neither should be treated as a toy.

Privacy

Do not type student essays, grades, photos, or names into any model, including a “harmless” image generator that could memorize a face. FERPA does not pause for a demo. Packets use public or fictional workplace data (generic yield numbers, a fictional lot ID, a public landscape). District-approved tools only. If students later try a tool at home, they should still withhold PII; this lesson models that withholding by never asking for it.

Reflection

  1. Which family do you over-trust because it looks finished?
  2. Where in an Idaho job you know would a simple tree or a line be more honest than a chatbot?
  3. What will you ask the next time someone says “the AI said”?

Homework

Family hunt (20–30 minutes, no new account). From a newspaper, a printed ad, a game, or a parent’s workplace story, find two systems that might use AI or data models. On paper, guess the family (text, image, class, number) and write five sentences each: purpose, possible training data, and one misuse. If you cannot tell, write what evidence you would need. Do not sign up for a tool to complete this.

Closing

Large language models speak in text. They are one family, not the whole field. If you can match a job to a family—and refuse the families that do not belong—you will not be stuck asking a chatbot to do a sensor’s work, a tree’s work, or a camera’s work.

Extensions and cross-curricular links

Go further

  • Block period: students build a paper decision tree for a local yes/no (game on or off for smoke; irrigation on or off) and contrast it with a three-sentence LLM-style blurb about the same decision.
  • Math extension: calculate a slope from the three yield points and discuss why extra points could change the line.
  • Art/CTE: label a generated image with a required caption (“not a photograph”) as if it were going on an Idaho tourism site.
Mathematics
Linear models and flowcharts are already in Algebra and logic units. This lesson names them as AI families with limits.
Science
Classification vs. quantitative prediction mirrors taxonomy vs. measurement. Students practice not mixing those jobs.
ELA
LLMs produce fluent text. Fluency is not evidence. Students compare a generated permit letter with a primary source.
CTE
Ag yield numbers, semiconductor inspection, and trades diagnostics use different model families. Brand chatbots are not the shop standard.