Building on Week 6
Week 6 addressed the misconception that AI systems understand users or act in their interest. The corrective was technical: recommendation systems maximize a reward signal, not user wellbeing; supervised classifiers learn patterns in labeled data, not truth.
Week 7 deepens those corrections. Neural networks raise the interpretability question that decision trees answered so conveniently — you could print a decision tree on a page and read its logic. A neural network with millions of parameters cannot be read. LLMs raise the fluency problem: a system that produces well-structured, confident text creates a far stronger illusion of understanding than a chatbot that clearly pattern-matches. Middle schoolers who use ChatGPT regularly have absorbed that illusion deeply. These pages offer tools for addressing it.
Neural Networks at the Middle School Level
The Perceptron as a Single Decision
The perceptron model from Topic 7a is accessible at the middle school level through a concrete framing: a perceptron is a single neuron-like unit that takes several inputs, weighs how important each one is, adds them up, and decides yes or no based on whether the total crosses a threshold. That is the entire mechanism.
A classroom analogy: imagine deciding whether to bring an umbrella. You check three things: is it cloudy (weight: 3), is rain predicted (weight: 5), is it summer (weight: 1). If the weighted total is above 6, you bring the umbrella. Below 6, you don't. A perceptron does the same thing — except the weights were not set by a person. They were adjusted thousands of times by a learning algorithm that compared the perceptron's outputs to the correct answers.
The Black Box Tradeoff
In Week 6, decision trees were praised partly for their transparency. Neural networks are the counter-example: they are far more capable but far less interpretable. A network with millions of parameters can recognize faces, translate language, and detect tumors in medical images — none of which is achievable by any decision tree you could print on a page. But you cannot read a neural network the way you can read a decision tree. You cannot follow its logic. You can only observe its outputs.
This tradeoff — capability versus interpretability — is one of the central tensions in modern AI. Middle schoolers who understand it are better equipped to ask the right questions when they encounter AI systems in the world: not just "does it work?" but "can anyone explain why it does what it does?"
An Activity: The Spam Filter
Email spam filters are a neural network application that middle schoolers already interact with, even if they don't know it. Present students with a simple scenario: an email system has learned to filter spam by being shown thousands of labeled examples (spam / not spam). Ask them: what features do they think the system learned to look for? What kinds of spam might it miss? What legitimate email might it accidentally block?
This activity builds intuition for how supervised learning generalizes — and fails to generalize — without requiring any mathematics. The discussion of false positives and false negatives connects directly to the bias discussions from Week 6.
Large Language Models at the Middle School Level
The Deepest Misconception to Address
Middle schoolers who use ChatGPT regularly have formed a strong intuition: the tool understands their question, searches for the answer, and reports what it found. None of those three things is accurate. LLMs predict the next token based on patterns in training data. They do not search. They do not retrieve. They do not know what is true.
The ELIZA story from Week 6 is still the best corrective for the "understanding" part. But LLMs present a harder version of the same problem: they produce outputs far more convincing than ELIZA, in far more domains, with far greater fluency. The mechanism is the same — pattern matching on prior input — but the scale makes the illusion nearly irresistible.
Why Hallucination Happens
LLMs hallucinate because they are trained to produce plausible continuations, not true ones. When asked a question, the model generates text that looks like the kind of answer that would appear after that question — because it has seen many question-answer pairs. If the correct answer is not well-represented in the training data, or if the question has no reliable answer, the model still produces confident text. It cannot distinguish between "I have good information about this" and "I have no information about this but I can generate something that sounds like an answer."
A useful middle school framing: imagine a very well-read student who has never been tested on anything. They have read a lot about history, science, and literature. But they have also read a lot of fiction, misinformation, and outdated sources — and they cannot tell which is which because they have never had to check. When you ask them a question, they give you the most plausible-sounding answer they can. Usually it is right. Sometimes it is confidently wrong.
A Classroom Activity: Spot the Hallucination
Give students a short AI-generated passage on a topic they know something about — local history, a book they have read, a science topic from class. Ask them to identify any claims that seem off and to verify them using a reliable source. Then discuss: how did the passage feel to read? Did it seem credible? What made the errors hard to spot?
This activity builds critical evaluation habits that transfer to every context where students encounter AI-generated text — which is increasingly everywhere.
Connections to the Broader 6–8 CS Curriculum
- Impacts of computing: The interpretability tradeoff and the hallucination problem both connect to CSTA standards on the social impacts of computing. Neural networks that cannot be explained raise accountability questions; LLMs that produce confident misinformation raise information integrity questions. Week 7 gives those standards concrete technical grounding.
- Data and analysis: Neural networks trained on biased data amplify that bias at scale — a direct extension of the data quality discussions from Week 5 and the bias discussions from Week 6.
- Digital citizenship: "Evaluate AI-generated content before trusting it" is more actionable than "think critically about what you read online" — and understanding why LLMs hallucinate makes the advice make sense rather than sounding like a generic warning.