Building on Week 6
Week 6 addressed the scope misconception: high school students who have used ChatGPT often treat LLMs as synonymous with AI, and everything else as either a precursor or a footnote. The Week 6 content — search, decision trees, the learning taxonomy — was a direct corrective. Students who worked through it now understand that LLMs are one application of a much larger field.
Week 7 lets those students go deep on the applications they actually care about. Neural networks and LLMs are where high school CS instruction can be most substantive, most technically honest, and most directly relevant to decisions students will face as citizens, workers, and potential technologists. The pages below are organized around what is within reach at 9–12 and what makes the biggest instructional difference.
Neural Networks at the High School Level
Perceptrons and Backpropagation: What Depth is Appropriate?
High school students can engage with the perceptron model at full technical depth: inputs, weights, weighted sum, threshold, output — and the training loop of forward pass, error measurement, weight adjustment, and repetition. They can trace a perceptron calculation by hand, reason about what a positive versus negative weight means (excite versus inhibit), and understand why adding layers changes what the network can learn.
Backpropagation is the mechanism by which error signals propagate backward through the network to adjust weights. High school students do not need the calculus formulation — but they can understand the intuition: each weight is adjusted in proportion to how much it contributed to the error. The network gets slightly better at the examples it just saw. Repeated across millions of examples, this produces a network that generalizes to new inputs. Students in AP CS A or a data structures course can engage with gradient descent as a concept even without the formal machinery.
The Interpretability Tradeoff as a Central Theme
The contrast between decision trees and neural networks is one of the most productive instructional threads in high school AI. Decision trees are interpretable but limited. Neural networks are powerful but opaque. This is not a flaw to be engineered away — it is a fundamental tension in the field.
High school students can reason about where this tradeoff matters most. A spam filter that occasionally misclassifies a legitimate email is a nuisance; an opaque neural network making parole decisions or loan approvals is a civil rights concern. The interpretability requirement is not absolute — it depends on the stakes of the decision, who bears the consequences of errors, and whether there is any avenue for challenging the outcome. Students who understand this can engage seriously with the AI ethics cases they will encounter in the broader world.
Deep Learning and Representation
The key insight of deep learning — that adding hidden layers allows the network to learn increasingly abstract representations — is accessible at high school with the right framing. Early layers in an image-recognition network learn to detect edges; later layers detect shapes; later still, objects. The network is not told to do this — it emerges from training. This is what makes deep learning qualitatively different from earlier machine learning approaches: the features are learned, not engineered.
Students in a data science or machine learning elective can explore this concretely using visualization tools that show what different layers of a trained network have learned to respond to. These tools make the abstraction hierarchy visible and give students a genuine sense of what "representation learning" means.
Large Language Models at the High School Level
The Architecture at a Level That Matters
High school students are ready for the full three-phase account of LLM development: pretraining on unlabeled text using next-token prediction, supervised fine-tuning on human-written examples of good behavior, and reinforcement learning from human feedback (RLHF) to align outputs with human preferences. They do not need the transformer architecture in detail — but they should understand that pretraining is where the model learns language, fine-tuning is where it learns to be helpful and safe, and RLHF is where it learns what humans tend to prefer.
That three-phase picture is directly relevant to the limitations discussion. Hallucinations arise primarily from pretraining: the model predicts plausible continuations based on statistical patterns in training data, with no mechanism for verifying factual accuracy. Fine-tuning can reduce obvious hallucinations but cannot eliminate them because the underlying mechanism — next-token prediction on noisy training data — remains. Students who understand this can reason about why the problem is systematic rather than correctable.
Handling the Confidence That Students Bring
High school students who use LLMs daily often arrive with one of two extreme positions: either the tools are transformatively powerful and the concerns are overblown, or they are unreliable and the concerns are understated. Both positions contain truth. The instructional goal is calibration.
LLMs are genuinely good at certain tasks: generating first drafts of structured text, explaining concepts at an accessible level, translating between languages, summarizing long documents, brainstorming options. They are systematically bad at certain other tasks: reliable factual retrieval, mathematical reasoning, tasks that require knowing what is actually true versus what sounds like it could be true. Students who can identify which category a task falls into — and who can check outputs against reliable sources when the stakes require it — are positioned to use these tools well.
LLMs as a Pedagogical Case Study
For students heading toward data science, software engineering, or any field that will involve building or deploying AI systems, LLMs are an ideal case study because they combine nearly every issue the course has addressed: bias baked into training data, reward signal misalignment (RLHF optimizes for human preference, not factual accuracy), the interpretability problem at extreme scale, and the gap between capability and reliability. A student who can reason clearly about LLMs has developed the analytical vocabulary to reason about any AI system they will encounter.
Connecting to AP Computer Science Principles
AP CSP's "Data" and "Impacts of Computing" big ideas both intersect directly with Week 7 content. The concept of training data and how its properties shape model behavior is a foundational AP CSP idea that Weeks 6 and 7 make technically concrete. LLM hallucination, bias amplification, and the interpretability tradeoff are all "impacts of computing" questions that require the technical background Week 7 provides in order to be answered substantively rather than rhetorically. Students who leave this course understand not just that these issues exist but why they arise from specific design choices.
Connections to the Broader 9–12 CS Curriculum
- AP Computer Science Principles: The data, algorithms, and impacts big ideas all intersect with Weeks 6 and 7. Decision trees, neural networks, LLM architecture, and the bias/interpretability discussions are all AP CSP-relevant at depth. Students preparing for the exam who understand these concepts have a significant advantage on the "impacts of computing" questions.
- Data science electives: Week 7 provides the conceptual foundations that data science electives often assume without teaching: what a neural network is, what training means, what the difference between supervised and unsupervised learning is, and why deep learning changed what was possible. Students who arrive at a data science elective with this background can engage immediately with implementation rather than spending weeks on conceptual orientation.
- Ethics and societal impacts: The SEC scenarios this week — reward signal misalignment, biased training data, hallucination in educational contexts — are the cases that appear in engineering ethics courses at the university level. High school is not too early for these conversations, and Weeks 6 and 7's technical foundations make them substantive rather than merely rhetorical.