Teaching "Privacy & AI" to Architect mode (ages 15–18)

Hook & Warm-Up

This age band's hook assumes real vocabulary and sets a technical tone from the first line. Read it as-is, then use it to preview the two intellectual threads (technical and regulatory) the lesson is going to weave together.

"Every app you use collects data. AI makes that data incredibly valuable — and vulnerable. Let's examine the technical and regulatory dimensions."

Frame the session explicitly: "Today isn't 'be careful online' advice — you've had that lesson many times already. This is about the actual mechanisms: what technical approaches exist to protect data while still letting AI use it, what legal frameworks exist to govern that, and where the real, unresolved tension sits between AI capability and privacy protection." If any students are considering law, policy, CS, or data science, note that this exact tension — how much data an AI system needs versus how much privacy a person is owed — is a live, unsettled question in both fields right now, not a solved textbook problem.

Opening question to set up the lesson's central claim: "Is there a technical fix that eliminates the privacy cost of AI entirely, while keeping all of AI's usefulness?" Let students argue both sides briefly, then be direct: no — every technique covered today (federated learning, differential privacy, homomorphic encryption) reduces or reshapes the tradeoff, but none eliminates it. That honesty upfront prevents the lesson from being received as "here's the solution" when it's really "here's how the tradeoff gets managed."

Main Activity

The in-app scenes are the same three as every age band — What AI Knows, How Data Spreads, Protect Yourself — but treat each as a doorway into one of the specific technical or legal concepts this age band's own quiz requires: federated learning, contextual integrity, homomorphic encryption, and the "data as labor" framework.

Scene 1 — What AI Knows: from raw signals to inference

Establish precisely what's technically happening: location, voice, and face data are raw signals; the privacy risk isn't really the raw signal itself but the inference an AI model can draw from aggregating many signals over time — a single GPS point is nearly harmless, but a model trained on months of them can infer home address, workplace or school, religious practice, health conditions (frequency of visits to a specific type of clinic), and social relationships (who you're regularly co-located with). This inference gap — between what was explicitly shared and what can be derived — is the crux of most modern privacy debate, more than the collection of any single data point in isolation.

Scene 2 — How Data Spreads: contextual integrity as a real framework

Introduce contextual integrity, a privacy theory developed by NYU philosopher and privacy scholar Helen Nissenbaum, precisely: it argues that privacy violations occur not simply when information flows or is shared, but specifically when information flows violate the context-specific norms that were supposed to govern that data's movement. A concrete example: sharing your grades with a teacher is normal within the context of school; that same teacher selling your grades to an advertiser is a privacy violation — not because the information moved, but because it moved in a way that violates the norms of the context it originated in. This reframes "privacy" away from simple secrecy ("is it hidden?") toward appropriateness of flow ("does this movement of information match the norms of the context it came from?") — a genuinely more useful lens for analyzing cases like a fitness app sharing health data with insurers, or a school app sharing behavioral data with a marketing partner.

Connect this directly to data brokers and tracking cookies: both are, under this framework, violations of contextual integrity almost by definition — data given in one context (browsing a specific website, using a specific app) flows into an entirely different context (a comprehensive advertising profile, or a stranger's dataset) that the original disclosure never implied consent to.

It's worth noting explicitly why this framework matters for AI specifically, beyond privacy debates in general: AI systems are unusually good at generating new context-crossing flows on their own, without a human deciding to move the data. A model trained to predict one thing (say, ad relevance) can incidentally learn to infer something entirely different and more sensitive (say, a likely health condition) purely as a side effect of finding whatever patterns are statistically useful — a flow no one explicitly designed or consented to, and one contextual integrity is well-suited to naming as a violation even when no single human decision caused it.

Scene 3 — Protect Yourself: the real technical toolkit

Cover three genuine technical mitigations with precision, since the quiz tests exact mechanisms rather than vague familiarity:

• Federated learning: trains a shared AI model across many decentralized devices without centralizing the raw underlying data — each device computes a local model update using its own private data, and only that update (not the raw data itself) is sent to a central server and aggregated. This is how, for example, a keyboard's next-word-prediction model can improve from millions of users' typing patterns without any company ever collecting a searchable copy of what anyone actually typed.
• Homomorphic encryption: allows computation directly on encrypted data, producing an encrypted result that, when decrypted, matches what you'd have gotten by computing on the plaintext directly — meaningfully different from ordinary encryption, which must be decrypted before any computation can happen on it. It remains computationally expensive relative to plaintext computation, which is why it's used selectively (specific high-sensitivity computations) rather than universally, as of current practice.
• The "data as labor" framework: a policy proposal, most associated with economist Glen Weyl and computer scientist Jaron Lanier, arguing that since user-generated data creates substantial economic value for platforms, users should be compensated for the data they generate — reframing data not as a free byproduct of using a service, but as a form of uncompensated labor. It remains a proposal, not implemented law anywhere at scale, and is genuinely contested among economists and technologists.

Have students complete the lesson individually, then discuss in pairs: "Of these three approaches (federated learning, homomorphic encryption, data-as-labor), which addresses a technical problem, and which addresses an economic/policy problem? Are they solving the same issue in different ways, or genuinely different issues?" (Expected answer: federated learning and homomorphic encryption are technical approaches to reducing data exposure; data-as-labor is an economic/policy reframing that doesn't reduce collection at all — it argues for compensating it instead. This distinction is the kind of thing that separates a surface-level answer from one that actually understood the lesson.)

Worth flagging as an honest limitation, since a sharp student may raise it independently: federated learning reduces raw data centralization, but the aggregated model itself is not automatically private — research on machine learning has documented "membership inference" and related attacks where an adversary can sometimes determine whether a specific individual's data was used in training a model, even without ever seeing the raw data directly. This doesn't invalidate federated learning as a real, useful mitigation; it does mean "federated" and "private" are not strictly synonymous, and serious deployments typically combine federated learning with additional protections like differential privacy rather than treating either alone as sufficient.

Discussion

• "Contextual integrity argues privacy violations are about information flowing outside its original context, not simply about secrecy. Does that framework change how you'd judge whether a specific real app's data-sharing practice is a violation?"
• "Federated learning avoids centralizing raw data, but the resulting model itself can still leak information about training data under certain attacks. Does that mean federated learning 'solves' the privacy problem, or just relocates it?"
• "If homomorphic encryption is currently too computationally expensive for most everyday uses, what does that imply about who currently benefits from strong technical privacy protections — and is that itself a policy problem?"
• "Evaluate the 'data as labor' framework: what would change, practically, if users were paid for the data they generate? What new problems might that create?"
• "GDPR and the Philippines' Data Privacy Act both grant rights to access, correct, and delete data. Contextual integrity suggests a different kind of protection — governing appropriate flow, not just access and deletion. Which do you think matters more for AI systems specifically, and why?"
• "Is there a real difference between a government using AI-derived data for surveillance versus a private company doing the same for commercial profiling? Should the law treat them differently, and does it currently?"
• "Federated learning doesn't guarantee a trained model can't leak information about individuals in its training data. Does that make it a meaningful privacy protection, a marketing claim, or something in between — and how would you argue for your position?"

Push for precision at this age — a strong answer names the specific mechanism (contextual integrity, federated learning, a specific legal right) rather than defaulting to general statements like "privacy is important" or "companies shouldn't do that."

Quiz Walkthrough

This age band's quiz override tests exact mechanisms and named frameworks — each question includes at least one plausible-sounding distractor, so the reasoning matters more than recognizing a keyword.

Wrap-Up & Extension

Close with the lesson's own summary line: "AI privacy combines technical measures like differential privacy, legal frameworks like GDPR, and ethical principles. Federated learning trains models without centralizing data. The tension between AI capability and privacy protection is fundamental." Add: "That word 'fundamental' is doing real work — this isn't a lesson that ends with a clean answer, because there isn't one yet. What you should leave with is the vocabulary to reason about the tradeoff precisely, instead of falling back on either 'privacy doesn't matter' or 'all data collection is theft.'"

Extension activity: Have students research one real, publicly documented case where a company's data practice was challenged under GDPR, CCPA, or the Philippines' Data Privacy Act (regulatory decisions, fines, and settlements are usually publicly reported). Have them summarize, in a short write-up, what specific data practice was challenged, which legal right or provision it allegedly violated, and how contextual integrity would independently analyze the same case — does the legal outcome match what contextual integrity would predict, or diverge from it? Require students to cite their source and distinguish clearly between what a source states directly and what they're inferring.

For a class with more time or a policy/debate component, run a structured debate: "Should the Philippines adopt GDPR-equivalent AI-specific privacy regulation, including rights against automated profiling?" Require both sides to engage with at least one specific mechanism from today's lesson (contextual integrity, federated learning's real limits, or the data-as-labor framework) rather than general privacy sentiment, and to address at least one genuine cost or tradeoff of their own position — stronger regulation has real compliance costs, especially for smaller Philippine startups and edtech companies, and that tradeoff deserves to be argued honestly rather than dismissed.