How Generative AI Works: A Lesson Plan for Grade 7

A student holds up their phone during your warm-up and says, “But isn’t ChatGPT just Googling faster?” You pause. The rest of the class is watching. You know the answer is no — but can you explain why, right now, in terms a 7th grader will actually understand? That moment, and the student’s genuine curiosity behind it, is the entire reason this lesson exists. By the end of this post you will have a word-for-word 50-minute plan that turns that question into a structured investigation — and sends students out with a mental model they will use for the rest of their lives online.

TL;DR: Generative AI is a very fast autocomplete. When a student types a question into ChatGPT, the model does not look up an answer — it predicts, one word at a time, which word is most likely to come next, based on billions of examples it read during training. That is why it sounds confident even when it is wrong. A tool trained to sound right will not always be right. This lesson helps grade 7 students see that distinction clearly, using their own phone’s autocomplete as the starting point.

What Is Generative AI, Actually?

Here is the version you can put on a slide and read aloud.

ChatGPT does not look up answers in a database. It does not search the web in real time and return a result. What it does is predict. Given every word typed so far, the model calculates which word is statistically most likely to come next — and places that word. Then it does it again for the next word. And again. Thousands of times per response, in fractions of a second.

The model learned to do this by reading an enormous amount of human text — web pages, books, forum threads, articles — and being trained to predict, accurately, what word tends to follow what word in that kind of text. Georgetown’s Center for Security and Emerging Technology describes this as “next-word prediction at scale”: the same process your phone keyboard uses to suggest the next word in a text message, but trained on billions of examples instead of your personal message history. (Georgetown CSET — The Surprising Power of Next Word Prediction)

The critical implication: the model was trained to predict plausible text, not to verify facts. When it produces a confident-sounding sentence, that confidence is a byproduct of pattern-matching — not a signal of accuracy.

The 50-Minute Lesson Sequence

This plan runs in one standard class period. No devices required for the first 20 minutes — students use their own phones only during the autocomplete hook. The investigation portion can use a class set of devices, a teacher-projected demo, or a single shared laptop per group.

Minutes 0–5 — Autocomplete Hook

Students discover that prediction is already built into the devices in their pockets.

Ask everyone to open their phone’s text-message app (or use the projected teacher phone if your school is 1:1 restricted). Have them type “I love” in the message field — not send, just type. Then: what three words does the phone suggest next?

Collect five responses on the board. You will get a range: “I love pizza,” “I love you,” “I love dogs.” Ask: why does your phone suggest a different word than your neighbor’s phone? Students typically arrive at the answer themselves within two minutes — the phone learned from your messages. Write that sentence on the board and leave it there. You’ll return to it.

Minutes 5–20 — From Phone to ChatGPT

Same idea. Vastly bigger training set.

Now point to the board. “Your phone autocomplete learned from your message history — maybe a few thousand texts. ChatGPT’s autocomplete learned from billions of texts from across the internet.” Draw this as a simple scale: one person’s texts on one end, the internet on the other. Same mechanism. Different scale.

Before anyone opens a device, distribute the fill-in-the-blank prediction sheet (or project it). Give students five sentence stems — “The capital of France is ___,” “Water boils at ___,” “The author of Romeo and Juliet is ___” — and ask them to predict what word comes next, the same way ChatGPT would: not by looking it up, but by gut instinct based on what they’ve read before.

After two minutes, ask: “Were you right? How did you know?” Students will say things like “I just knew,” “it sounded right,” or “I’ve seen it before.” That is the moment. “That’s exactly what ChatGPT does — except it has seen it billions of times more than you have. So it sounds even more sure.” Check their predictions against ChatGPT live, projected. Compare the AI’s wording with the students’ wording. Often they are nearly identical — because both are pattern-matching from the same cultural corpus.

Minutes 20–35 — Small-Group Investigation

Students run the autocomplete mechanism into a real mistake.

Divide the class into groups of three or four. Each group gets a set of two or three factual questions — mix topics: history, science, current events, pop culture. Groups type each question into ChatGPT (or the teacher types and projects while they observe), screenshot or write down the output, then spend five minutes attempting to verify one answer using a second source.

The questions that produce the most productive discussion are ones where ChatGPT’s statistically likely answer turns out to be wrong or outdated: “Who is the current president of [a country]?”, “What is the tallest building in the world?”, “What year did [a recent event] happen?” When the verification fails, students have found the autocomplete mechanism producing a confident mistake. Have each group flag: did ChatGPT predict its way into something wrong?

Minutes 35–45 — Class Discussion: Why the Confident Mistake

Name the causal chain together, out loud.

Bring the class back. Go around the room: what did each group find? Then walk through the mechanism on the board, step by step:

1. The model predicts the most statistically likely next word.
2. It does not check whether that word is true.
3. When the most likely word happens to be wrong, the model places it with the same confidence as a right answer.
4. There is no internal “I’m not sure” flag. It does not have one.

Ask students: “So when should you trust what ChatGPT says?” The productive answer is not “never” — it is “verify anything that matters, the same way you would verify any single source.” This keeps the lesson from becoming anti-technology panic and positions students as informed evaluators rather than passive users.

Minutes 45–50 — Exit Ticket

Two sentences. That’s the whole deliverable.

The prompt: “In your own words — why does AI sound sure even when it’s wrong? Write 2 sentences.”

A strong response names the mechanism: “ChatGPT predicts the next word based on patterns, not facts. It doesn’t know what’s true — it knows what usually comes next.” Any student who can write those two sentences has the mental model. That’s your formative check.

Why Does AI Hallucinate? The Mechanism, Not the Mystery

A grade 7 class that runs the investigation above will typically find at least one confident wrong answer within the first five minutes of the small-group work. This is not a bug the developers missed. It is a direct consequence of what the model was trained to do.

The causal chain is short: the model learns to predict plausible text → it places the statistically likely word → the statistically likely word is occasionally, factually wrong → the model has no mechanism to notice the difference → it delivers the wrong word with the same fluency as the right word. MIT’s explainer on generative AI frames this as the core tension: the system was built to produce coherent text, not to verify whether that text is correct. (MIT News — Explained: Generative AI)

This directly addresses ISTE 1.3.d — students build knowledge by actively exploring real-world issues and pursuing investigation. The investigation here is not abstract. Students run a query, verify the output, and find the seam between “plausible” and “true.” That is real epistemic work, not a definition exercise.

It also connects to AI4K12 Big Idea #4: AI can impact society. A model that confidently predicts false information — at the scale of millions of users per day — has societal implications that go well beyond a wrong homework answer. The courtroom hallucination cases, the misinformed medical queries, the viral pizza-glue incident: all of them are the same mechanism, in different contexts.

For a deeper classroom treatment of specific hallucination examples students can verify live, see the companion post: AI hallucination examples for grade 7 ELA: 5 to teach Monday.

How to Adapt This for ELA and Science

The core lesson above is content-neutral — it runs in any classroom with a projector and five minutes of phone access. Two subject-specific angles extend it further.

For ELA Teachers

CCSS.ELA-LITERACY.RI.7.8 asks students to evaluate the reasoning and evidence in informational text, distinguishing claims supported by evidence from those that are not. ChatGPT output is informational text. Students who have run the investigation in this lesson are primed to read any AI-generated paragraph the way they read a primary source: where is the claim? what evidence supports it? what would disprove it?

Use the exit ticket responses as a class text set. Students read each other’s two-sentence explanations, evaluate which ones accurately name the mechanism, and discuss what a persuasive but wrong explanation looks like — the same skill as evaluating a flawed argument in a textbook excerpt. Zero new prep. Same reading standard, new text type.

For Science Teachers

Generative AI is applied statistics. The model assigns probabilities to candidate next-words and selects from the high-probability end of the distribution. That is a probability problem.

Your grade 7 probability unit likely covers sample space and likelihood. The autocomplete analogy maps directly: a larger training corpus narrows the probability distribution around the correct answer — the same way a larger sample size in an experiment narrows the margin of error. Students who see generative AI as “statistics on text” have a concrete application for probability concepts that otherwise feel abstract. The connection is real, not forced.

Standards at a Glance

Quick-scan for the teacher checking coverage before the lesson goes on the plan book:

• ISTE 1.3.d — Knowledge Constructor: build knowledge by actively exploring real-world issues and pursuing answers through active investigation
• AI4K12 Big Idea #4 — AI Can Impact Society: understanding how AI affects individuals, communities, and decision-making at scale
• CCSS.ELA-LITERACY.RI.7.8 — evaluate the reasoning and evidence in informational text; distinguish claims supported by reasons and evidence from claims that are not
• ISTE 1.2.b — Digital Citizen: students recognize the rights, responsibilities, and opportunities of living and learning in an interconnected digital world
• AI4K12 Big Idea #3 — Representation and Reasoning: how AI systems learn from data and what that means for their outputs

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-Grab Resources

If you want to run this lesson without building every piece yourself, two printables do the heavy lifting.

The Generative AI Deep-Dive Lesson — Next-Word Prediction + Hallucinations, Grades 6-8 is the printable version of this exact lesson sequence: the fill-in-the-blank prediction activity, the small-group investigation sheet, the exit ticket, and the teacher guide — all formatted for immediate classroom use. No slide deck to assemble.

Once students understand how generative AI predicts text, a natural follow-up question is: what kinds of AI systems exist beyond chatbots? The Types of AI Deep-Dive Lesson — Narrow vs General vs Super AI, Grades 6-8 is the companion lesson that extends the conversation into the full landscape of AI systems — from the narrow AI in a spam filter to the general AI students keep reading about in the news.

For a free starting point before committing to the deep-dive, the free AI literacy lesson plan for grade 7 covers the foundational “what is AI” period and pairs well as the session before this one.

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