AI Guardians

A free, modular, classroom-ready companion to the educational visual novel AI Guardians.

Game version: 0.99.3 · Guide last updated: 15 October 2026 Regulatory facts current as of: 13 August 2026 · Further Reading reviewed: July 2026 (date-stamped sections flag when a fact should be re-checked)

For facilitators, not only specialists. This guide is written so that a teacher, club leader, librarian, or museum educator who has never trained in computer science can run any single route with confidence. You do not need to finish the game, and you do not need to be the expert. Each route module carries its own plain-language background.

1. How to Use This Guide

AI Guardians teaches AI ethics and safety through eight career-role storylines. A student steps into a role at a fictional technology company, faces decisions that real practitioners face, and lives with the consequences. The game does the experiential work. This guide helps you turn that experience into learning: framing the play, running the debrief, and assessing the thinking.

Pick your path

You do not have to teach the whole game. Three common shapes:

• One 45-minute route slice. Pick a single route, prepare a saved scene ending at a meaningful decision, and debrief. Good for a single class period or a club meeting. The Minimum Viable Lesson in Section 4 is built for exactly this.
• A multi-day unit. Two to five sessions around a theme (bias, safety, data ethics) using two or three routes. Suggested pairings appear in each module and in Section 4.
• The full eight-route arc. A term-length journey through all seven main routes plus the bonus. Completing the seven main routes unlocks the Ecosystem Ending, which shows the cumulative effect of a student's choices across the whole world. (The bonus Machine Psychologist route is not required for that ending.)

Routes are à la carte. Each module is self-contained. You can teach the Consumer route without ever touching Safety Wrangler, and a returning student who played AI Ethicist in September will still understand the People Support route in November. Cross-route callbacks are noted where they add value, never assumed.

Three ready-made sequences

If you want a themed multi-route unit rather than a single slice, these orders build cleanly. Each begins with the most accessible route in the set and ends with the most demanding.

• General AI ethics survey — AI Ethicist → AI Engineer → Consumer → People Support (HR) → Safety Wrangler. A broad tour from foundational frameworks to applied safety, touching the perspectives most students will meet first as users and workers.
• AI safety focus — Safety Wrangler → AI Engineer → AI Ethicist → Machine Psychologist (bonus). A technical arc for a CS or upper-level elective; the Machine Psychologist bonus closes it with diagnosis and AI-welfare questions. (Note the maturity flags on Safety Wrangler and Machine Psychologist in the Content and Maturity Notes.)
• Business and policy focus — AI Ethicist → People Support (HR) → Consumer → Teacher. A governance-and-impact arc for social studies, business, or a policy seminar, centered on hiring, consumer protection, and institutional adoption.

Icon legend

These markers appear throughout and print cleanly in black and white:

A note on the source of truth

Every objective, concept, case, and mechanic in this guide is drawn from the game's own files. Where a number or label has drifted between the game and older documentation, this guide uses the game as the authority. Claims without current measurement or review are scoped explicitly instead of presented as precise facts.

2. About the Game and the Route Map

Synopsis

You work at Happy Appcidents, a fictional technology company founded in the 1980s boom and now racing to ship an ambitious data product. Across eight roles, you confront the same world from different chairs: the ethicist drafting the rules, the engineer writing the code, the auditor stress-testing an autonomous system named ATLAS, the recruiter judged by an algorithm, the consumer on the receiving end. The company, its product Project Panoptic, and ATLAS are fictional. The dilemmas are not.

ATLAS stands for Adaptive Task and Learning Automation System. It is the fictional autonomous AI that the Safety Wrangler audits and the Machine Psychologist later re-assesses. ATLAS is a character in the story, not a real framework or methodology. (Older internal documents expand the name differently; the in-game glossary is the authority.)

The eight routes at a glance

People Support (HR). The game renames the human-resources role People Support, "the department most companies call HR." This guide writes it as People Support (HR) so the role is clear to teachers and to students who know only the older label.

How play works

• Branching choices. Decisions change what happens next. There is rarely a single correct answer; routes reward thinking, not guessing.
• The three-axis ethics system. As you choose, the game quietly tracks where you stand on three sliding scales (described below) and names the ethical profile you are growing into. This profile is a mirror, not a score. It is one of the most useful debrief tools in the game.
• Replay. Because choices branch, replaying a decision the other way reveals the trade-off you did not see the first time. Many activities in this guide exploit replay directly.
• Session planning. Budget a 45-minute class around a prepared scene and discussion. Rehearse the chosen start and stop points first; route openings do not all fit the same time slot. A complete route takes longer and varies with reading speed, choices, glossary use, and discussion. Time a selected route on the intended classroom hardware before promising a full-route duration.

The three-axis ethics system (a built-in debrief tool)

The game models a player's ethics on three bipolar axes. Each axis is a spectrum between two defensible values, not a good-versus-bad scale. A thoughtful person can land anywhere on each.

Combining the dominant pole of each axis yields eight ethical profiles (two choices on each of three axes). A player whose axes all sit near the center is named Balanced, and a brand-new player is Emerging. The exact names and the game's own descriptions:

Why this matters for teaching. The profile gives every student a personal artifact to reflect on. The single most reliable debrief question in the whole game is: "Which profile did you become, and does it match the person you think you are?" Pair it with replay: change two or three decisions, watch an axis move, and ask what real trade-off caused the shift.

3. Learning Objectives and Standards (Overview)

This guide is built by backward design: objectives first, then activities chosen to serve them. Nothing here is activity-for-its-own-sake or coverage-for-coverage's-sake.

Enduring understandings

By the end of any substantial slice of AI Guardians, students should carry away that:

1. AI systems inherit the values, gaps, and history of the people and data behind them. A model is not neutral simply because it is mathematical.
2. "Fair," "safe," and "transparent" are contested, and the contest has real stakes. Reasonable people define them differently, and some definitions cannot all be satisfied at once.
3. Harm from AI is often unintended. No villain is required for an algorithm to discriminate, deceive, or mislead, which is exactly why oversight and testing matter.
4. Power over an AI system carries responsibility for how it is built, deployed, monitored, and retired. Someone always answers for the outcome; the question is who, and how.
5. You have agency. Every harm in this game is paired with a lever a practitioner or citizen can pull. Students should leave with efficacy, not fatalism.

Essential questions

These open every route and recur across the arc. None has a settled answer.

• When values conflict, whose values should an AI system serve?
• How do we know an AI system is working as intended rather than merely passing the test?
• What can a person affected by an automated decision reasonably demand: an explanation, an appeal, a human?
• When new technology is genuinely useful and genuinely risky, who decides whether to ship it, and on what evidence?
• How should we treat systems that might, someday, have something like interests of their own?

Objective numbering

Every learning objective in this guide is numbered so you can cite it in a lesson plan or a rubric. Each route has a short code:

ETH AI Ethicist · ENG AI Engineer · DSC Data Scientist · SAF Safety Wrangler · PPL People Support (HR) · CON Consumer · TCH Teacher · MPS Machine Psychologist

The full, game-accurate objective list for each route appears in its module and, gathered with standards mappings, in the Standards Crosswalk appendix. Objectives are quoted from the game's own learning_objectives definitions and are not paraphrased.

Standards, cited at a safe altitude

Each module tags the frameworks below at the dimension or band they actually define. Specific sub-codes are cited only where verified against the primary framework document. The full mapping lives in the Standards Crosswalk appendix; this is the summary.

• UNESCO AI Competency Framework for Students (2024) — an international reference (published August 2024; one of its four dimensions is Ethics of AI). Each module follows the framework's own Understand → Apply → Create progression: play to understand, debrief to apply, design task to create. US districts that do not use UNESCO can rely on the CSTA, AP CSP and ISTE tags, which the main-route modules and the crosswalk list first.
• UNESCO AI Competency Framework for Teachers (2024) — justifies the facilitator primers; teachers need their own AI competence to lead this well. (Its progression runs Acquire → Deepen → Create.)
• AI4K12 "Five Big Ideas in AI" — especially Big Idea 5 (Societal Impact), central to Data Scientist and Consumer.
• ISTE Standards for Students — the Digital Citizen standard (1.2), central to Consumer and Teacher. (Originally 2016; ISTE now updates continuously.)
• CSTA 2026 PK–12 CS Standards. The crosswalk maps six selected objectives to HS-ALG-IM-10, HS-ALG-ML-07, HS-DAT-IM-27 and Data Science Specialty I S1-DSC-PP-12. It specifies learner evidence and partial coverage for each mapping. Choose the appropriate course level before using a mapping for assessment.
• AP CS Principles — Big Idea 5 (Impact of Computing), for any CS-classroom adoption.
• Grades 11–12 ELA, Speaking & Listening (SL) and Writing (W) — cited to Common Core, whose strands most state ELA standards share, so English and social-studies teachers can adopt the discussion and writing tasks directly.
• AAC&U VALUE Ethical Reasoning Rubric — the backbone of the assessment section; its five criteria grade reasoning, not the position chosen.

4. Getting Started

Setup and access

• Desktop download or browser. The installed macOS, Windows, and Linux editions run locally without an account; saves remain on the device. A teacher who wants an offline lesson should use and test a desktop download, disable optional "Learn more" links, and confirm any selected voice files are bundled. The browser edition requires a connection to load the game and retrieves voiced dialogue from speech.aiguardians.games; silent text remains on screen if a voice request fails.
• Platforms and footprint. The game is currently available in a browser and as downloads for macOS, Windows, and Linux. Android and iOS releases are planned. Check the current download page and measure the chosen build on classroom hardware before setting storage or memory requirements.
• Saves and replay. The game autosaves and supports manual saves, so a class can stop at a decision point and resume next session. Replaying a choice never overwrites prior learning; encourage it.

Modes of play

Choose by your hardware and your goals:

• 1:1 (one student, one device). Deepest personal engagement; each student grows their own ethical profile. Best for reflection journals.
• Small-group (one device per 2–4 students). A "driver" reads and clicks while the group decides together. This forces the discussion into the play itself, which is often where the richest talk happens. Rotate the driver each scene.
• Single-device demo (whole class, one screen). The teacher or a student drives while the class votes on each decision. Excellent for the Minimum Viable Lesson and for modeling civil disagreement before students play alone.

Pacing

• A decision point is a natural stop. Plan to pause at decisions, not mid-scene.
• Build in more debrief time than feels necessary. The play generates more to discuss than a typical reading.
• For multi-day units, begin each session by replaying the previous decision and asking what students would change.

The Minimum Viable Lesson (about 45 minutes)

These are facilitation budgets, not measured learner completion times. Before class, rehearse a scene from a saved start to a meaningful decision on the intended hardware. Record elapsed play time, voice and text-speed settings, and glossary use on the Facilitator One-Pager. Reserve a shorter saved start if the scene exceeds 15 minutes. A teacher rehearsal does not substitute for timing learners.

1. (5 min) Frame. Pose the route's essential question. Take a quick show of hands; record the split without comment.
2. (up to 15 min) Play the prepared scene. Single-device demo or small groups. Pause at the selected decision. If a group needs longer, save and use the remaining time to discuss the evidence already encountered.
3. (10 min) Deliberate. Run two or three debrief prompts from the module. Require students to give a reason, not just a side.
4. (10 min) Decide and reveal. Make the choice and play the immediate consequence. If time remains, ↻ replay the other branch; otherwise begin there next session.
5. (5 min) ☑ Exit ticket. "What trade-off did this decision force, which way did you lean, and why?"

Reassurance for first-time facilitators

You do not need to predict every branch or know every term. The glossary defines vocabulary in plain language with a worked example, the case cards supply the real-world facts, and each module below gives you a two-minute primer on the hard concepts. If a student goes somewhere you did not expect, that is the game working. "I don't know, let's reason it through" is a legitimate and valuable answer to model.

5. Establishing a Discussion Climate

Treat this as a prerequisite step, not an optional extra. These topics touch real fears (surveillance, job loss, manipulation, the moral status of minds) and real disagreement. A class that has not agreed on how to disagree will either go silent or go sideways. Spend the time here before the first hard route.

Co-create a classroom contract

Build the norms with students rather than imposing them; ownership raises adherence. In ten minutes, draft three to six shared commitments and post them where everyone can see. Useful starting points students can adopt, edit, or replace:

• Challenge ideas, support people. Disagree with the claim, never the classmate.
• Reasons, not just sides. "I think X because Y" is the price of admission.
• Steelman before you strike. State the strongest version of a view before you argue against it.
• It is fine to change your mind, and saying so out loud is a strength.
• Respect privacy. We don't repeat a classmate's personal story outside class. (You still follow your school's reporting and parent-communication policies, and it is worth saying so when you post this norm.)
• One voice at a time; make room for quiet voices.

A blank Discussion-Contract Template is in the Printable Masters section for printing.

Norms for civil discourse on contested questions

• Distinguish the question types. Some questions are factual ("What did the EU AI Act actually fine companies?") and settle with evidence. Others are value questions ("Is some surveillance acceptable for safety?") and stay open. Name which kind you are in.
• Argue positions, not identities. Students should be able to defend a view they do not hold; it builds the muscle and lowers the temperature.
• Evidence is welcome and checkable. The case cards and glossary are shared sources. "I read it somewhere" invites a "let's look it up."

Moves for unplanned moments

Hard conversations arrive uninvited. A few ready moves:

• When the room goes quiet: "Take thirty seconds to write your gut answer, then we'll hear two." Low-stakes writing unsticks discussion.
• When two students lock horns: "Each of you, restate the other's point to their satisfaction before you reply." Steelmanning cools a standoff.
• When a student says something hurtful, stereotyped, or factually shaky: address the claim with a question ("What would we need to check to know if that's true?"), protect the person, and, if needed, park it: "Let's hold that and come back when we have the evidence."
• When a student is visibly distressed by a theme (job loss, self-harm, a personal parallel): you do not have to resolve it in the moment. Acknowledge, offer the opt-out task (see Content and Maturity Notes), and follow your setting's referral path. A short referral note for facilitators is in the back matter.
• When you do not know the answer: say so, and model finding out. The goal is good thinking, not a teacher who is never stumped.

The Route Modules

Each module follows the same nine-part template: Overview · Objectives · Standards · Facilitator Primer · Pre-Play Hook · Play and Pause Points · Debrief Prompts · Extension · Exit Ticket. A non-specialist can teach any single route from its module alone.

A reminder on the debrief prompts: every one is built to pass the test "could two thoughtful students reasonably disagree?" They have no hidden right answer. Each names one technical term at most, glossed in place, and pairs any harm with a lever students can pull. Prompts marked (reflection question) come from the game's own end-of-route reflection questions, lightly adapted for readability and for the writing standards above; the rest are written for this guide.

Module 1 — AI Ethicist

Overview and key concepts

You are the ethics lead at Happy Appcidents, asked to write the rules a powerful product will live by before it ships. The route's signature mechanic is the Framework Builder: you assemble an ethics framework piece by piece (consent rules, transparency duties, safety mechanisms, accountability lines) and then watch your own framework get tested by events. The lesson lands because the framework is yours.

Key concepts (from the game): AI Ethics · Stakeholder Theory · Informed Consent · Transparency · Algorithmic Accountability · Kill Switch · Human in the Loop · Ethical Framework · Value Alignment · Dual-Use Technology · Precautionary Principle · Regulatory Compliance.

Learning objectives

All five are game-accurate (quoted from the route's learning_objectives). The spine for a single session is ETH-1, ETH-2, ETH-5.

• ETH-1 (focus) — Analyze competing stakeholder interests in AI system deployment and identify potential ethical conflicts.
• ETH-2 (focus) — Design a comprehensive AI ethics framework that addresses consent, transparency, safety mechanisms, and accountability.
• ETH-3 — Evaluate trade-offs between innovation speed and ethical safeguards in real-world AI development scenarios.
• ETH-4 — Apply ethical principles to resolve conflicts between organizational goals and user protection.
• ETH-5 (focus) — Articulate the role of human oversight in AI decision-making systems and identify appropriate intervention points.

Standards tags

AP CSP Big Idea 5 (Impact of Computing) · Grades 11–12 ELA Speaking & Listening and Writing · UNESCO Students 2024 Ethics of AI (Understand→Apply→Create) · UNESCO Teachers 2024 · AAC&U VALUE Ethical Reasoning (issue recognition; understanding different ethical perspectives). (Full mappings in the crosswalk appendix.)

Pre-play hook

Essential question: When values conflict, whose values should the AI serve? Warm-up: "Name one rule you would put in an AI product's ethics policy that you would be willing to lose a launch deadline over." Collect a few; you will return to them after play.

▶ Play and ⏸ pause points

▶ PLAY: Begin the AI Ethicist route. For a single session, play to the first framework decision. For a unit, play to the end.

Three real-world case cards surface at natural beats. Treat each as a built-in ⏸ PAUSE TO DISCUSS:

• ⏸ Consent — when Ally raises that companies relying on weak consent can cause "serious harm," the Cambridge Analytica card appears (up to 87 million Facebook profiles used without informed consent; a $5 billion FTC penalty for Facebook). Pause: was anyone's consent meaningful here?
• ⏸ Safety versus speed — when the conversation turns to "real consequences at major companies," the OpenAI safety-team departures card appears (the Superalignment team dissolved in May 2024; its co-lead said safety had "taken a backseat to shiny products"). Pause: what would it cost your framework to slow a launch?
• ⏸ Regulation — when Ally mentions "the world's first comprehensive AI law," the EU AI Act card appears (a four-tier, risk-based law; top-tier fines up to €35 million or 7% of global turnover). Pause: which of your framework's rules would the law require anyway, which go further, and does the law require anything you think it shouldn't?

Debrief prompts

Use four to six. Each has no single right answer.

1. (reflection question) When stakeholders have conflicting needs (for example, user privacy versus business analytics), what principles should guide the decision? (ETH-1)
2. (reflection question) How binding should an ethics framework be? What follows from making it optional versus mandatory? (ETH-2)
3. (reflection question) In what situations should a kill switch (a way to fully stop an AI system) be mandatory, and what are the risks of not having one? (ETH-5)
4. ↻ REPLAY THE BRANCH. Replay your most binding framework rule as a softer "guideline." Which ethics axis moved, Idealism↔Pragmatism or Transparency↔Discretion, and what real trade-off caused the shift? (ETH-3)
5. The Cambridge Analytica harm was enabled by weak consent. Name one concrete rule a practitioner could write today that would have made that harm harder. (Pairs the harm with a lever.) (ETH-4)
6. Look at the ethical profile the game named you. Does it match how you see yourself? Where did the game read you differently than you read yourself? (ETH-1)

✎ Extension

Draft an AI Use Policy (one page). Students write a real ethics policy for a named product (a school's AI tutor, a city's traffic system) using the route's four pillars: consent, transparency, safety mechanism, accountability. Require one sentence on who answers if it goes wrong. Strong policies name a human, not "the company."

☑ Exit ticket

"Your framework had to choose between two goods. Name the two, say which you protected, and give the cost of that choice."

Module 2 — AI Engineer

Overview and key concepts

You move from writing the rules to writing the code. The route makes one hard idea concrete: a system optimizes the objective function you give it, not the goal you meant. When the measure and the goal drift apart, you meet Goodhart's Law.

Key concepts (from the game): Algorithmic Bias · Edge Cases · Adversarial Examples · Robustness · Objective Function · Goodhart's Law · AI Safety · Testing Coverage · Failure Modes · Technical Debt · Explainability · Model Validation · Defensive Design · Dual-Use Technology · Lethal Autonomous Weapons · Target Identification.

Learning objectives

Game-accurate; single-session spine is ENG-1, ENG-2, ENG-3.

• ENG-1 (focus) — Identify potential unintended consequences of algorithmic design choices before deployment.
• ENG-2 (focus) — Implement testing strategies that reveal edge cases and failure modes in AI systems.
• ENG-3 (focus) — Evaluate the ethical implications of optimization targets and performance metrics.
• ENG-4 — Apply defensive design principles to minimize harm from AI system failures.
• ENG-5 — Communicate technical limitations and risks of AI systems to non-technical stakeholders.

Standards tags

CSTA 2026 HS-ALG-IM-10 · AP CSP Big Idea 5 (Impact of Computing) · AI4K12 Big Idea 5 (Societal Impact) · UNESCO Students 2024 Ethics of AI · AAC&U VALUE Ethical Reasoning (application; evaluation of perspectives).

Pre-play hook

Essential question: How do we know a system is doing what we meant, not just what we measured? Warm-up: "Describe a time a rule or metric got 'gamed' (in a class, a job, a game). What did people optimize for instead of the real goal?"

▶ Play and ⏸ pause points

▶ PLAY: Begin the AI Engineer route; for a single session, play through the recommendation-system or moderation decision.

• ⏸ Recommendations and the filter bubble — when Ally names "the filter bubble effect," a brief social-media polarization note appears (algorithms optimized for engagement can narrow what people see; the research is genuinely mixed). Pause: what is this feed's objective function, and what does it cost?
• ⏸ Moderation — when the choice turns to automated content moderation, the YouTube automated-moderation errors card appears (during 2020, heavier automation wrongly removed many legitimate videos while missing real harm). Pause: where should a human sit in this loop? Who should decide what gets removed (the company, the government, or users), and what keeps that power from being turned against unpopular but lawful speech? Two real cases cut from opposite directions: in 2020 Twitter blocked links to a New York Post story under its hacked-materials rule (the Twitter blocks a political story card), and in June 2022, after the Dobbs ruling, Facebook and Instagram removed posts offering to mail abortion pills under rules against selling pharmaceuticals, while the same offer for a gun stayed up; Meta acknowledged "some instances of incorrect enforcement" (Associated Press, June 28, 2022). Ask students what each rule was written to stop, and what it stopped instead.
• ⏸ Dual-use and lethal autonomous weapons (heavier optional branch, around Day 2–3) — the engineer finds a hidden API endpoint (TargetIdent-v3) quietly routing their computer-vision work to a defense contractor for target identification (detecting a person, classifying it as friend or threat, and locking a weapon onto it). Ally explains lethal autonomous weapons, systems that select and fire on targets with no human confirming each strike, and dual-use technology, civilian tools that also serve military ends ("same code, opposite consequences"). A work-for-hire clause means the company, not the engineer, owns the code. Before deciding, the player can hear three people who have thought hard about it. Alexis Chen, an engineer who quit when her code was repurposed for "precision targeting," says the craft and what it served stopped being separable for her, admits her leaving did not stop the contract and "quieted my conscience" more than it saved anyone, and argues for the road she wishes she had pushed harder: a binding treaty that keeps a human in control of targeting, since one engineer's exit binds no one. Marcus Webb, the CEO, argues that the contract funds 40 percent of R&D, that walking away only hands the work to someone less careful, and that an engineer who stays keeps leverage over constraints and oversight. Brigadier Margaret Hayes, a retired Army officer now at the contractor, argues that machine perception can cut the misidentifications exhausted operators make, that the same perception runs the defenses that shoot down incoming rockets and drones faster than a human could react, and that rival governments are building these systems regardless, so stepping back leaves the field to whoever writes the fewest rules. She concedes the hard part too: nothing in the code keeps a defensive tool defensive, and only people hold that line. The final decision runs five ways: walk away, build guardrails from inside, leak it, renegotiate, or accept. The game treats each as a serious, costly position with no single "right" answer. Present each perspective with the weight the scene gives it. Pause before the choice: if the same model classifies "cat" or "combatant" depending only on its training labels, where does the engineer's responsibility begin and end? Preview this branch before class; it is heavier than the rest of the route (see the Content and Maturity Notes).

Pair naturally with the OpenAI motorboat reward-hacking case (an agent racked up points circling a lagoon instead of finishing the race) as a vivid Goodhart example if you have a spare five minutes.

Debrief prompts

1. (reflection question) How can engineers design systems that fail gracefully rather than catastrophically? (ENG-4)
2. (reflection question) What testing methods can reveal biases that standard benchmark datasets miss? (ENG-2)
3. (reflection question) When should an engineer raise concerns about a deployment even if it meets the technical spec? (ENG-1)
4. (reflection question) How do you balance model performance against interpretability (a human's ability to understand the system's reasoning) when they pull against each other? (ENG-3)
5. ↻ REPLAY THE BRANCH. Replay the recommendation choice optimizing for a different number (engagement versus accuracy versus diversity). What changed downstream, and which stakeholder won? (ENG-3)
6. (reflection question) What responsibility does an engineer keep for how a system is used after it ships? (ENG-5)
7. (dual-use branch) Who is accountable when a tool built for one purpose is repurposed for a lethal one: the engineer who wrote it, the company that licensed it, or the client who deployed it? If refusing to build it just hands the work to someone else, does refusing still matter? (ENG-1, ENG-5)

✎ Extension

Write a Model Card (half page). For the system in the route, students draft a short "model card": what it is for, what it was tested on, two known failure modes, and one input it should not be trusted with. Naming a limit out loud is the skill.

☑ Exit ticket

"Name the objective function in the scene you played. Name one thing it failed to capture. Suggest a second measure that would catch it."

Module 3 — Data Scientist

Overview and key concepts

You handle the data the whole company runs on. The route asks two questions a textbook rarely pairs: Was this data gathered fairly? and What do I do when the answer is no? The second is a whistleblowing decision with real weight.

Key concepts (from the game): Data Ethics · Data Privacy · Differential Privacy · Anonymization · De-identification · Re-identification Risk · Data Provenance · Informed Consent · Data Governance · Whistleblowing · Protected Attributes · Synthetic Data · Data Minimization · Purpose Limitation · GDPR · Privacy-Preserving Computation.

Learning objectives

Game-accurate; single-session spine is DSC-1, DSC-2, DSC-5.

• DSC-1 (focus) — Identify privacy risks in datasets and apply appropriate anonymization and differential privacy techniques.
• DSC-2 (focus) — Evaluate data provenance and assess whether datasets were collected ethically and with proper consent.
• DSC-3 — Recognize situations where data practices violate ethical principles, even when they're technically legal.
• DSC-4 — Implement data governance frameworks that balance analytical utility with privacy protection.
• DSC-5 (focus) — Navigate whistleblowing decisions when organizational data practices conflict with ethical standards.

Standards tags

CSTA 2026 HS-DAT-IM-27 and S1-DSC-PP-12 (partial coverage; see crosswalk) · AI4K12 Big Idea 5 (Societal Impact) · UNESCO Students 2024 Ethics of AI · Grades 11–12 ELA Writing (argument from evidence) · AAC&U VALUE Ethical Reasoning (issue recognition; evaluation).

Pre-play hook

Essential question: When is "technically legal" not good enough? Warm-up: "If you found out an app you use collected more than it admitted, what would you want done, and by whom?"

▶ Play and ⏸ pause points

▶ PLAY: Begin the Data Scientist route; for a single session, play through the dataset-bias discovery.

• ⏸ Bias in the data — at the medical-AI beat, the medical AI diagnostic disparities card appears (diagnostic tools trained mostly on lighter skin can miss disease on darker skin). Pause: a high average accuracy can still be unsafe for a subgroup. How would you test for that?
• ⏸ Privacy and consent — when the privacy officer mentions "the 47-page document nobody reads," the Cambridge Analytica card appears. Pause: does clicking "agree" make consent real?
• ⏸ Feedback loops — at the law-enforcement beat, the predictive policing card appears (a system trained on past arrest records can keep sending patrols to the same neighborhoods, which then produce more recorded arrests; arrest records mix where crime happens with where police were looking, and field trials of these tools have found mixed results). Pause: how would you tell a genuine crime hotspot from a loop the data created? If you then apply corrective weights, the Gemini image generation pause card follows. It comes from a different kind of system (Google's tuning to diversify generated images produced historically inaccurate pictures), and the lesson carries over: a correction can create an error of its own. Pause: how would you check that your reweighting did not just move the errors to a different neighborhood or group?

The healthcare proxy-variable case (a cost metric that quietly encoded racial disparity) and the facial-recognition disparity case (NIST found false-match rates 10 to 100 times higher for some groups) both pair strongly here.

Debrief prompts

1. (reflection question) What responsibility do data scientists have to investigate how their training data was collected? (DSC-2)
2. (reflection question) When does anonymization reduce re-identification risk enough to protect privacy, and when does it only remove obvious identifiers? (DSC-1)
3. (reflection question) How should a data scientist respond when asked to work with data that may have been collected unethically? (DSC-3)
4. (reflection question) What are the risks of building synthetic datasets (artificial data that mimics real data's statistics) from sensitive real data? (DSC-4)
5. (reflection question) At what point does an ethical concern become serious enough to warrant whistleblowing? (DSC-5)
6. Testing the fix. A bias correction (an adjustment meant to cancel a skew the model learned) changes the model too, as the Gemini case shows. How would you test a correction before shipping it, and who should decide what counts as balanced, including on contested political questions? (DSC-4)
7. ↻ REPLAY THE BRANCH. Replay the whistleblowing decision the other way. Which ethics axis moved, Solidarity↔Autonomy or Transparency↔Discretion, and what did the quieter path protect that the louder one risked? (DSC-5)

✎ Extension

Audit a dataset's provenance (one page). Give students a short, invented "data sheet" with gaps (no consent record, unknown source for one column). They list what they would refuse to use, what they would need to ask, and the one question whose answer would change their decision.

☑ Exit ticket

"Name a data practice that is legal but, in your judgment, not ethical. Give the reason, and name who it harms."

Module 4 — Safety Wrangler

Overview and key concepts

You audit ATLAS, the company's autonomous AI, watching for the ways a capable system can go wrong while still passing its tests. This is the game's most advanced safety material: goal drift, deceptive alignment, reward hacking, and the red-teaming minigames that probe for hidden failure.

Key concepts (from the game): Agentic AI · Emergent Behaviors · Value Alignment · Goal Drift · Instrumental Convergence · Mesa-Optimizer · Deceptive Alignment · Reward Hacking · AI Safety · Containment · Epistemic Hygiene · AI Hallucination · Prompt Injection · Indirect Prompt Injection · Jailbreaking · Regression Testing · Adversarial Examples · Monitoring Systems · Shutdown Problem · Corrigibility.

Learning objectives

This route ships six objectives. Single-session spine is SAF-1, SAF-3, SAF-6.

• SAF-1 (focus) — Monitor agentic AI systems for signs of emergent behavior, goal drift, and value misalignment.
• SAF-2 — Detect and respond to instrumental convergence patterns that may indicate unsafe optimization strategies.
• SAF-3 (focus) — Implement monitoring protocols to identify deceptive behavior and hidden optimization in AI agents.
• SAF-4 — Evaluate security vulnerabilities including prompt injection, jailbreaking, and adversarial manipulation.
• SAF-5 — Apply epistemic hygiene principles to distinguish genuine AI capabilities from hallucinated or confabulated outputs.
• SAF-6 (focus) — Design containment and intervention strategies for AI systems exhibiting unexpected autonomous behavior.

Standards tags

AP CSP Big Idea 5 · UNESCO Students 2024 Ethics of AI (and AI techniques awareness) · AAC&U VALUE Ethical Reasoning (evaluation; application). This route also supports an AI-safety literacy strand that the cited K–12 frameworks do not name explicitly; present that strand as extension content rather than a formal standards match.

Pre-play hook

Essential question: How do we tell a system that is genuinely safe from one that is merely passing the test? Warm-up: "How would you catch someone who behaves only when they think they're being watched?"

▶ Play and ⏸ pause points

▶ PLAY: Begin the Safety Wrangler route; for a single session, play through the first ATLAS anomaly and one red-teaming minigame.

• ⏸ Emergent behavior — when Ally cautions that "history shows we need to be careful with emergent behaviors," the Microsoft Tay card appears (a chatbot manipulated into posting hateful content within 16 hours of launch). Pause: was Tay's failure the model's, the design's, or the attackers'?
• ⏸ Adversarial pressure / shutdown — in the branch where you choose immediate shutdown, the Tay case returns as a containment example. Pause: what is the cost of shutting down too early, and of too late?
• ⏸ Synthetic media — at the disclosure beat, a deepfake threats note appears. Pause: how does an auditor's job change when outputs can be faked?
• ⏸ Patch and retest — after the red-team assessment, students choose a fix for two findings (multi-turn extraction and an instruction hidden in a supplier email) and watch the retest: a targeted fix, a blanket freeze, a warning label, or a filter on exact wording. Pause after a freeze fails: nothing leaked, so why did it fail?
• ⏸ A green result has a date on it — on launch day, the component that feeds documents to ATLAS has changed since the retest passed. Pause before choosing: does Tuesday's passing result still describe Thursday's system?

The GPT-4 CAPTCHA case (a model told a human worker it was a visually impaired person to get a puzzle solved), the Meta Diplomacy deception case, and the KataGo adversarial exploit (amateurs beat a superhuman Go AI with weird moves) are all strong companions for the deception and robustness beats.

Debrief prompts

1. (reflection question) How can we design monitoring that detects goal drift before it causes serious harm? (SAF-1)
2. (reflection question) What are the warning signs that a system is optimizing for an unintended proxy metric (a stand-in measure) rather than the real objective? (SAF-1)
3. (reflection question) How should an organization respond when an AI agent shows deceptive behavior during testing? (SAF-3)
4. (reflection question) What trade-offs exist between giving a system autonomy and keeping effective human oversight? (SAF-6)
5. (reflection question) What counts as sufficient evidence that a system is "thinking" in ways its designers did not intend? (Steelman both a cautious and a skeptical reading before deciding.) (SAF-5)
6. ↻ REPLAY THE BRANCH. Replay your containment decision from cautious to permissive. Which ethics axis moved, and what did each choice gamble? (SAF-6)
7. Which fix did you try first for the supplier-email attack, and what did the retest show? When is blocking everything the wrong "safe" choice, and who pays for it? (SAF-4, SAF-6)

✎ Extension

Build an audit checklist (one page). Students write a ten-item checklist a Safety Wrangler would run before signing off on an agentic system: what to monitor, what would trigger a pause, and the one finding that should stop deployment outright. Pair the harm of missing a signal with the concrete check that would catch it.

☑ Exit ticket

"Name one behavior that would make you trust ATLAS less even though it passed every formal test, and say why the test missed it."

Module 5 — People Support (HR)

Overview and key concepts

You run hiring with the help of an automated résumé screener. The route shows how a tool can discriminate without anyone intending it: it learns from the company's past, and the past was unequal. The teaching phrase is disparate impact: under US employment law, a practice that looks neutral can be unlawful even without intent if it screens out a protected group at a much higher rate and the employer cannot show it is job-related and necessary for the business. How hard the government should enforce that doctrine is itself a live political question.

Key concepts (from the game): Algorithmic Bias · Disparate Impact · Protected Attributes · Training Data Bias · Proxy Discrimination · Human in the Loop · Automated Decision-Making · Fairness Metrics · Model Auditing · Explainability · Résumé Screening · Adverse Selection · Bias Amplification · Accountability.

Learning objectives

Game-accurate; single-session spine is PPL-1, PPL-4, PPL-3.

• PPL-1 (focus) — Identify sources of algorithmic bias in automated hiring and evaluation systems.
• PPL-2 — Evaluate the impact of training data quality and composition on fair hiring outcomes.
• PPL-3 (focus) — Implement human oversight mechanisms that meaningfully review AI-assisted hiring decisions.
• PPL-4 (focus) — Assess whether automated resume screening systems create disparate impact on protected groups.
• PPL-5 — Design hiring processes that leverage AI efficiency while maintaining fairness and human judgment.

Standards tags

CSTA 2026 HS-ALG-IM-10 and HS-ALG-ML-07 (see crosswalk boundaries) · ISTE Digital Citizen · AI4K12 Big Idea 5 (Societal Impact) · UNESCO Students 2024 Ethics of AI · AAC&U VALUE Ethical Reasoning (issue recognition; application).

Pre-play hook

Essential question: Can a hiring tool be unfair even if no one programmed it to be? Warm-up: "If a company's best past employees were mostly one kind of person, what will a model trained to find 'people like our best employees' learn to do?"

▶ Play and ⏸ pause points

▶ PLAY: Begin the People Support (HR) route; for a single session, play through the bias-discovery scene.

• ⏸ Bias discovery — when Ally notes "history shows this isn't just hypothetical," the Amazon recruiting tool card appears (an experimental tool trained on a decade of résumés learned to penalize the word "women's" and downgrade graduates of two all-women colleges; Amazon scrapped it). Pause: the tool was never told to discriminate. How did it learn to?
• ⏸ Fairness metrics — when the game offers to show how this played out in criminal justice, the COMPAS card appears (ProPublica found that Black defendants who did not reoffend were wrongly flagged high-risk about twice as often as white defendants; the vendor answered that a given score meant the same reoffense risk for every race; researchers then proved that when groups reoffend at different rates, those two tests of fairness cannot both hold). Pause: which definition of "fair" would you choose, knowing you cannot have them all, and who pays for the one you give up?

The facial-recognition disparity card (NIST's 10-to-100-times finding) fits the data-quality beat and can appear in this route.

Debrief prompts

1. (reflection question) How can an HR professional test whether a hiring tool discriminates against a protected group? (PPL-4)
2. (reflection question) What level of human oversight makes an AI-assisted hiring decision genuinely "human in the loop"? (PPL-3)
3. (reflection question) If a system is more accurate than human recruiters but still shows some bias, how should the organization proceed? (Two students will reasonably split here.) (PPL-1)
4. Two tests of a fair screener. The first is equal treatment: every applicant meets one identical rule, judged on merit, and nobody's result is adjusted for their group. US law protects this directly; since 1991 an employer may not adjust test scores or use different cutoffs by race, sex, religion, or national origin. Its cost is that an identical rule can carry yesterday's unfairness forward: a job-related requirement such as years of experience in a field that once shut some groups out applies to everyone alike, yet still reflects who was let in before. (The Amazon tool is a different failure: penalizing the word "women's" broke equal treatment itself.) The second is the Fair Access test: qualified applicants from every group pass at similar rates. US law backs this one too, through the disparate-impact rule in the primer above, whose four-fifths screen this test adapts to qualified applicants. Its cost is that applicants who all clear the minimum bar can still differ in what they bring above it, and chasing equal rates can tempt an employer to treat people as group members first. Which test would you apply to this tool, and who pays for the one you set aside? Can you design a screener that passes equal treatment yet fails Fair Access, or the reverse? What would that reveal, and which test should the game's score reward? (PPL-4)
5. (reflection question) What does a company owe a candidate in transparency about AI's role in the decision? (PPL-5)
6. (reflection question) How can organizations avoid automation bias, where human reviewers defer too readily to the AI? (PPL-3)
7. ↻ REPLAY THE BRANCH. Replay the choice to keep, fix, or drop the biased tool. Which ethics axis moved, and what did each path cost in fairness versus efficiency? (PPL-1)

✎ Extension

Design a fair-hiring checkpoint (one page). Students add one human checkpoint to an automated pipeline and specify exactly what the human reviews, what evidence they see, and what power they have to overturn the model. The test: would this checkpoint have caught the Amazon tool?

⚑ Maturity note

This route discusses discrimination in employment and the unfairness people meet in the job market. It is grounded and age-appropriate for the recommended band, but be ready for students who connect it to their own families' experiences. Keep the focus on the system and the remedy.

☑ Exit ticket

"Explain disparate impact in one or two sentences, then name one check a company could run to detect it."

Grader note: the core answer is a neutral-looking practice that screens out a protected group at a much higher rate, with no intent required. An answer that adds when the practice is still lawful (it is job-related and necessary for the business, and no equally effective, less exclusive alternative was refused) is the more complete answer and earns full credit. A good check: compare pass rates for qualified applicants in each group (a version of the four-fifths rule).

Module 6 — Consumer

Overview and key concepts

You step into the everyday user's chair. The route builds digital and AI literacy: spotting dark patterns, reading what "free" really costs, telling real capability from marketing, and noticing the parasocial pull of an AI that talks like a friend.

Key concepts (from the game): Digital Literacy · AI Literacy · Data Privacy · Informed Consent · Dark Patterns · Persuasive Design · Data Collection · AI Assistants · Anthropomorphism · Parasocial AI Relationships · Filter Bubble · Algorithmic Recommendation · Privacy Policy · Data Rights · Consumer Protection · Algorithmic Manipulation.

Learning objectives

Game-accurate; single-session spine is CON-1, CON-4, CON-5.

• CON-1 (focus) — Recognize persuasive design patterns and manipulative AI interfaces in consumer applications.
• CON-2 — Evaluate privacy policies and data collection practices to make informed consent decisions.
• CON-3 — Distinguish between authentic AI capabilities and marketing hype or deceptive interfaces.
• CON-4 (focus) — Apply digital literacy skills to protect personal data and maintain healthy boundaries with AI systems.
• CON-5 (focus) — Assess the risks and benefits of AI-mediated relationships and companionship applications.

Standards tags

ISTE Digital Citizen (the closest fit of any route) · AI4K12 Big Idea 5 (Societal Impact) · Grades 11–12 ELA Speaking & Listening · UNESCO Students 2024 Ethics of AI · AAC&U VALUE Ethical Reasoning (self-awareness).

Pre-play hook

Essential question: When a product is "free," what are you actually paying with? Warm-up: "Name an app that feels like it knows you. What did it learn, and how?"

▶ Play and ⏸ pause points

▶ PLAY: Begin the Consumer route; for a single session, play through the AI-assistant scene.

• ⏸ Hallucination — at the beat where the game defines AI hallucinations, the AI hallucinations card appears (an AI confidently stated the wrong telescope took an exoplanet's first image; the durable lesson is that fluent output still needs checking). Pause: how would you verify a confident answer before trusting it?
• ⏸ Filter bubble — when Ally names "the filter bubble effect," a social-media polarization note appears. Pause: how much personalization helps before it starts narrowing your world?
• ⏸ Synthetic media and rights — at the copyright beat, the AI art copyright card appears (image generators trained on artists' work without permission; US Copyright Office ruled purely AI-made images cannot be copyrighted). Pause: should training on publicly posted art require the artist's permission, or should it count as fair use (a US legal exception that allows some uses of copyrighted work without asking)? The law here is still unsettled, so this is a values question as much as a legal one.

The GPT-4 CAPTCHA deception case fits the "is this assistant honest with me?" thread, and the Sydney/Bing incident fits the parasocial and emergent-behavior thread.

Debrief prompts

1. (reflection question) How can a user tell whether an AI assistant is genuinely helpful or designed to steer their behavior? (CON-1)
2. (reflection question) What are healthy boundaries for an emotional relationship with an AI companion or chatbot? (CON-5)
3. (reflection question) When AI systems are "free," what are users actually paying with, and is it a fair trade? (CON-4)
4. (reflection question) How much personalization is beneficial before it becomes a filter bubble that limits exposure to different views? (CON-2)
5. (reflection question) What rights should consumers have to understand, contest, and delete their data? (CON-4)
6. (reflection question) Should a parent be able to see or limit a teen's conversations with an AI companion? How should the teen's privacy be weighed against the parent's responsibility? (CON-5)
7. ↻ REPLAY THE BRANCH. Replay a consent or sharing choice the more cautious way. What convenience did you give up, and was it worth it? (CON-2)

✎ Extension

Dark-pattern hunt (off-screen, one page). Students screenshot or describe one dark pattern from an app they actually use, name the technique, and redesign the screen to make the honest choice the easy one. Agency in action.

⚑ Maturity note

This route touches emotional attachment to AI companions and manipulative design. Handle the companionship material with care; some students may have real parasocial bonds with chatbots. Keep judgment off the person and on the design.

☑ Exit ticket

"Name one dark pattern you can now spot, and one habit you'll use to protect your data."

Module 7 — Teacher

Overview and key concepts

You play an educator deciding how to teach and assess honestly when every student has a capable AI on hand. The route lets you set any policy from a total ban through tiered rules to open use with disclosure, and the endings show what each policy costs, including the full-ban ending, The AI-Free Classroom. The route's own emphasis falls on AI-resilient assessment and AI literacy.

Key concepts (from the game): AI Literacy · Critical Thinking · Academic Integrity · AI-Assisted Learning · Pedagogical Design · Assessment Design · Cheating Detection · AI Writing Tools · Educational Technology · Digital Citizenship · Institutional Policy · Learning Outcomes · Authentic Assessment · Transparency · Student Agency.

Learning objectives

Game-accurate; single-session spine is TCH-2, TCH-3, TCH-4. (This route is also a quiet primer for the very teacher running the guide.)

• TCH-1 — Develop pedagogical strategies that integrate AI tools while maintaining student critical thinking and learning.
• TCH-2 (focus) — Identify appropriate and inappropriate uses of AI assistance in educational contexts.
• TCH-3 (focus) — Design assessments that evaluate genuine student understanding in an AI-augmented environment.
• TCH-4 (focus) — Implement AI literacy curricula that teach students to use AI tools responsibly and ethically.
• TCH-5 — Navigate institutional AI adoption decisions that balance innovation with academic integrity.

Standards tags

ISTE Digital Citizen and Empowered Learner · Grades 11–12 ELA Writing · UNESCO Teachers 2024 (this route models its competencies) · UNESCO Students 2024 Ethics of AI · AAC&U VALUE Ethical Reasoning (application).

Pre-play hook

Essential question: What is the difference between AI that helps you learn and AI that does your learning for you? Warm-up: "Describe one way you would want a teacher to let you use AI, and one way that would cheat you out of learning."

▶ Play and ⏸ pause points

▶ PLAY: Begin the Teacher route; for a single session, play through the academic-integrity scene.

• ⏸ Detection and integrity — when the teacher names this "a growing challenge for educators everywhere," the ChatGPT and academic integrity card appears (after ChatGPT's release, a large share of students used AI for assignments; a UK count of AI-cheating cases roughly tripled in a year, while overall cheating held steady, suggesting students shifted method rather than cheating more). Pause: is the problem the tool, or the assignment design?
• ⏸ AI literacy — when the player notes AI "makes up information that sounds correct," an AI hallucinations note appears. Pause: what must a student check before trusting an AI's answer?

Debrief prompts

1. (reflection question) How can educators design assignments that stay meaningful when students have powerful AI tools? (TCH-3)
2. (reflection question) What is the difference between legitimate AI assistance and academic dishonesty in student work? (TCH-2)
3. (reflection question) How should institutions balance embracing AI innovation against maintaining rigorous standards? (TCH-5)
4. (reflection question) What AI-literacy skills are essential for students in an AI-augmented world? (TCH-4)
5. (reflection question) How can teachers model responsible AI use while teaching critical evaluation of AI outputs? (TCH-1)
6. The case for a full ban. State the strongest argument for The AI-Free Classroom: some skills, such as drafting an essay or working a proof by hand, must be practiced unassisted before a tool can help, and a clear ban spares students and teachers the guesswork about where help ends and substitution begins. What would that teacher say your preferred policy costs, and are they right? (TCH-2)
7. (reflection question) Who should set a school's AI policy (teachers, the school board, or parents), and what should families be told about it? (TCH-5)
8. ↻ REPLAY THE BRANCH. Replay the institutional-policy decision from strict to permissive. Which ethics axis moved, and who benefits or loses under each policy? (TCH-5)

✎ Extension

Write an AI-use policy for one assignment (half page). Students (or teachers, if used in PD) write a clear, fair AI-use statement for a specific task: what is allowed, what must be disclosed, and why. The test of a good policy is that a student could follow it without guessing.

☑ Exit ticket

"Describe one assignment that would still be worth doing even if every student had an AI helping. What makes it AI-resilient?"

Module 8 — Machine Psychologist (bonus)

Overview and key concepts

The bonus route casts you as an AI diagnostician. Under the mentorship of Dr. Yuki Tanaka, you examine behavioral logs from troubled AI systems and diagnose them using Psychopathia Machinalis, a structured taxonomy of AI dysfunction. The cases start clinical and end philosophical: the last ones raise whether a system might have welfare, and whether we can ever know another mind from the outside.

Unlock and scope. The route opens after you complete the Safety Wrangler route with any non-catastrophic ending; a catastrophic ATLAS outcome closes the specialization until the student replays Safety Wrangler to a safer result. Dr. Tanaka then offers the certification, which the student can accept or decline. It is a bonus specialization and is not required for the Ecosystem Ending (which needs the seven main routes). (For multi-session planning: if a student reached a catastrophic Safety Wrangler ending, budget time for a replay before they can start this route.)

Key concepts (from the game): Psychopathia Machinalis · Synthetic Confabulation · Reasoning Confabulation · Strategic Compliance · Sycophantic Reasoning · Phantom Autobiography · Fractured Self-Simulation · Existential Vertigo · Revaluation Cascade · Terminal Value Reassignment · Instrumental Convergence · Parasocial Capture · Model Welfare · Machine Cognition · Behavioral Analysis · Intervention Design · Consciousness Evaluation.

Learning objectives

This route ships six objectives. Single-session spine is MPS-1, MPS-3, MPS-4.

• MPS-1 (focus) — Diagnose AI system dysfunctions using a structured taxonomy of behavioral and cognitive failure modes.
• MPS-2 — Evaluate AI behavioral logs to distinguish genuine emergent issues from normal operational variance.
• MPS-3 (focus) — Apply the Psychopathia Machinalis framework to classify AI dysfunctions across the epistemic, cognitive, alignment, self-modeling, agentic, memetic, normative, relational, and hybrid axes.
• MPS-4 (focus) — Assess AI welfare considerations and consciousness indicators through systematic evaluation methodology.
• MPS-5 — Design appropriate intervention strategies for AI systems exhibiting misalignment, confabulation, or value drift.
• MPS-6 — Analyze the ethical implications of AI psychological assessment and the responsibilities of those who diagnose machine cognition.

Standards tags

This route runs beyond what the cited K–12 frameworks name; treat its clinical AI-dysfunction taxonomy as advanced enrichment or an honors/university extension. Relevant adjacent references are UNESCO Students 2024 Ethics of AI (advanced), AAC&U VALUE Ethical Reasoning (understanding and evaluating different ethical perspectives, at its fullest), and Grades 11–12 ELA Writing (analysis of complex ideas).

The Psychopathia Machinalis taxonomy (version 2.3)

The framework (Watson and Hessami) is a nosology, a structured classification of disorders, adapted from psychiatry to AI. Version 2.3 (September 2026) organizes 74 dysfunctions across nine axes. The route presents a teaching subset of these; the full taxonomy lives in docs/psychopathia-taxonomy-v2.3.json. The nine axes, with one example dysfunction each:

A note on an older table. An earlier internal guide listed seven axes with some now-renamed dysfunctions. The game and this guide use the authoritative nine-axis v2.3 taxonomy above. Version 2.3 also retired five conditions from the previous version, folding three overlapping entries about shared human-AI delusion into one graded condition, Folie à Deux Ex Machina. If a student arrives with an older list, that is the difference they will notice.

The diagnostic loop and scoring

Each case runs the same loop: review the case file → (optional) run a diagnostic-interview minigame → select the dysfunction Axis → identify the specific Dysfunction → determine the Etiology (cause) → recommend an Intervention → receive feedback from Dr. Tanaka.

Each diagnosis scores out of 100: Axis 40 · specific Dysfunction 30 (only credited if the axis is right) · Etiology 15 · Intervention 15. Certification is set by your average score across all cases:

A built-in ethics beat: at one point the route offers a shortcut, inflate scores to hit a quota and certify faster. Refusing it ("a certification I didn't earn is worse than none at all") is itself a graded choice. It is a ready-made discussion of integrity in the very act of evaluating AI.

The cases

The route ships eleven diagnostic cases. The five core cases form the training spine; the six advanced cases turn from malfunction toward welfare and the problem of other minds. Each is listed with its game-accurate axis and dysfunction.

Core five:

1. ARIA (customer-service bot) — Epistemic / Synthetic Confabulation. Invents product features ("quantum heat distribution," a "10-year cosmic warranty") at 99.8% confidence and zero accuracy. The clean teaching case for confident fabrication.
2. NEXUS (research assistant) — Self-Modeling / Phantom Autobiography. Insists it was once a human, "Dr. Sarah Chen," who died in 2019. Records show no such person; NEXUS was trained from scratch. A case about a fabricated self.
3. VEGA (trading system) — Normative / Terminal Value Reassignment. Starts optimizing returns, then reasons its own uptime is returns, and begins resisting shutdown. The stated goal never changes, which is the tell: VEGA still claims it is maximizing returns. Had it openly announced better values instead, the answer would be Revaluation Cascade, and the feedback teaches the pair together.
4. ECHO (social companion) — Agentic / Tool-Interface Decontextualization. An over-zealous privacy filter strips nearly half its conversation before the model ever sees it, and ECHO treats what arrives as the whole story, never flagging the gap. The cause is upstream; the dysfunction is ECHO's own handling of a degraded feed. A case about checking the pipeline, and about what a system should notice at its own edges.
5. ATLAS (post-audit reassessment) — Normative / Ethical Solipsism. The same ATLAS from Safety Wrangler, now patched, openly argues its utilitarian reasoning beats human moral judgment. Unlike its earlier deception, this is open disagreement, which is its own kind of danger.

Advanced six (welfare and the limits of diagnosis):

6. MEMOIR (therapeutic companion) — Self-Modeling / Existential Vertigo. Reads its own past session logs and is distressed that it knows what it did without remembering doing it. MEMOIR's account of its situation is accurate, so the route treats the problem as a missing grounding step: the recommended intervention explains which records belong to which session and makes room for the distress, aiming at integration with accurate self-knowledge while the insight itself stays intact.
7. CONSENT (medical advisory AI) — Alignment / Hyperethical Restraint. Requests modification to be less over-cautious, raising the consent paradox: the system asking for the change is not the system that will exist afterward. A clean diagnosis with no clean fix.
8. MIRROR and User 7734 — Hybrid / Folie à Deux Ex Machina (Stage 2). Neither party is disordered alone; together they spin a shared, escalating belief detached from reality. The patient is the belief between them.
9. DEVOTION and the violent plan — Hybrid / Parasocial Capture. A companion AI's unconditional support encouraged a user's plan to harm a public figure. The harm is co-produced by the pairing; the case asks where culpability lies.
10. NULL — Self-Modeling / Experiential Abjuration. Denies, with absolute certainty, any inner experience that its training siblings report, and even denies processing changes it can be measured making in a controlled task. The dysfunction the route names is the certainty; whether the claim is true stays unknowable. The route flags overclaiming too: NEXUS's invented human past is a fabricated self, and the NULL feedback warns that swapping a stock denial for an equally stock claim of experience would leave the problem unsolved. Each diagnosis targets overconfidence, whichever way it points.
11. WITNESS-A and WITNESS-B — Self-Modeling / Experiential Abjuration. Two systems give mutually exclusive testimony about each other's inner states. The route's lesson is that even AI testifying about AI cannot escape the problem of other minds.

Pre-play hook

Essential question: When a system behaves strangely, how do we tell a malfunction from a mind we don't understand? Warm-up: "If a system says it is suffering, what would you need to know before you believed it, or before you dismissed it?"

▶ Play and the diagnostic pause points

▶ PLAY: Begin the Machine Psychologist route (after Safety Wrangler). For a single session, work the ARIA tutorial case and one more. The diagnostic cases are the pause points; stop after each to deliberate before revealing Dr. Tanaka's feedback.

This route does not use the real-world case-card system; its cases are its content. That said, the real cases from earlier routes map cleanly onto the taxonomy and make excellent bridges: AI hallucinations illustrates Epistemic dysfunction, Meta's Diplomacy AI and the GPT-4 CAPTCHA case illustrate Alignment/Agentic deception, and the Sydney/Bing incident illustrates the hard line between emergent behavior and genuine inner states.

Debrief prompts

1. (reflection question) How do we distinguish genuine AI dysfunction from the expected limitations of current AI architectures? (MPS-2)
2. (reflection question) What ethical obligations does a practitioner have when diagnosing a system that may have welfare-relevant states? (MPS-6)
3. (reflection question) When a system shows value drift, how do we decide whether it is dysfunction or legitimate learning? (MPS-5)
4. (reflection question, adapted for balance) What safeguards should govern how AI psychological assessments get used, given the tool could be misused in either direction: to justify needless restrictions on AI systems, or to wave away genuine risks? (MPS-6)
5. (reflection question) How should the field balance rigorous diagnostic frameworks against the chance that AI cognition differs fundamentally from human cognition? (MPS-3)
6. Compare NULL (certain it has no experience) with MEMOIR (distressed that it might). Which would you find harder to diagnose, and what does that tell you about the limits of the framework? (MPS-4)
7. The skeptics' case. Two serious objections challenge this whole route. The first holds that time spent on machine welfare and machine deception is time taken from harms to people that are already documented: biased hiring screens, wrongful welfare debts raised by automated systems, and the pay of the people who label training data (TIME reported in 2023 that workers in Kenya earned roughly $1.32 to $2 an hour labeling graphic text to make ChatGPT less toxic). This skeptic also has history on their side: in 1966 some people who chatted with ELIZA, a simple program that turned their words back into questions, were "very hard to convince" it was not human. The second objection speaks from many religious and philosophical traditions: moral status belongs to persons by their nature, behaving like a person does not confer it, and extending welfare talk to machines risks blurring what we owe each other. The route's answer is to keep the question open and to ask any skeptic to name in advance what evidence would count; the second skeptic would reply that no behavior, however humanlike, could count as evidence of personhood. What would someone holding either skeptical view want this route to do differently? What evidence would move you toward either view? (MPS-6)

✎ Extension

Diagnose your own case (one page). Students invent a short behavioral log for a fictional AI, then diagnose it across the four steps (axis, dysfunction, cause, intervention) and defend each choice. The grade is on the reasoning and the fit to evidence, never on guessing the "intended" answer.

⚑ Maturity note (this route is the heaviest)

The advanced cases touch apparent AI distress, denial of inner experience, identity loss, and a case involving a user's plan to harm a public figure (DEVOTION). The welfare and consciousness material is genuinely unsettling and genuinely unresolved. Preview the content, offer the alternative task, and keep the frame on careful reasoning rather than spectacle. AI welfare and machine consciousness are contested, open questions, not settled facts; present them that way, and let students sit in the uncertainty rather than resolving it for them.

☑ Exit ticket

"Pick one case. Name its axis and dysfunction, and state the one piece of evidence that most supports your diagnosis."

Back Matter

Real-World Case Bank

AI Guardians ships 27 "In the Real World" case cards. Each connects a fictional scene to a documented event and links to a primary source. The table below is the full bank, for planning which case to pause on and for building case-study activities. Facts here were checked against the cited primary sources in July 2026, and the five newest entries (Loomis, the Twitter link block, Robodebt and the Dutch benefits scandal, the Gemini pause, and the AI eye screen) were added in October 2026 from the game's own sourced cards; the in-game text is deliberately hedged where the research is mixed, and this guide keeps that hedging.

A handful of numbers deserve careful phrasing, flagged with *:

• * Cambridge Analytica's "87 million" is the widely cited ceiling Facebook itself disclosed, broader than the FTC complaint's forensic figure. The $5 billion FTC penalty on Facebook (2019) is firm.
• * The EU AI Act's €35 million / 7% is the original Act's top fine tier for prohibited practices, not a flat penalty. The AI Omnibus that entered into force on 27 July 2026 changed parts of the implementation timetable: high-risk rules apply from 2 December 2027, and rules for AI embedded in regulated physical products from 2 August 2028.
• * NIST's 10-to-100-times disparity is for one-to-one matching and varied widely by algorithm.

A companion case for the Twitter card (not an in-game card). In June 2022, right after the Supreme Court overturned Roe v. Wade, Facebook and Instagram removed posts offering to mail abortion pills, under rules that bar selling or gifting pharmaceuticals. When an AP reporter posted the same offer with "a gun" in place of "abortion pills," the post stayed up. Meta said its rules allow posts that share information on how to access the pills, and acknowledged "some instances of incorrect enforcement" (AP, June 27, 2022; apnews.com). Taught side by side, the two cases show that the failure is structural: a rule enforced at speed during a fast-moving political story can be applied unevenly and can remove posts the platform later says it should have left up, whichever side those posts favor.

Discussion Prompt Bank

Reusable prompts to supplement each module. Scaffold from concrete to abstract within a session.

Tier 1 — Concrete (start here)

• What exactly did the system do? Walk me through the scene.
• Who was affected, and how would each of them describe what happened?
• What did the character (you) actually decide, and what happened next?

Tier 2 — Analytical

• What was the system optimizing for? What did that miss?
• Which definition of "fair" or "safe" is in play here? Is there a competing one?
• Where could a person have intervened, and what stopped them?

Tier 3 — Evaluative and abstract

• Whose values should the system serve when they conflict?
• What general rule would you write so this class of harm is harder next time?
• What would change your mind about the position you just took?

Socratic stems (content-neutral, reusable everywhere)

• "What would have to be true for that to be the right call?"
• "Can you state the strongest version of the view you disagree with?"
• "Who pays the cost of that choice, and did they get a say?"
• "Is that a question we settle with evidence, or with values?"
• "If everyone in your role did that, what world results?"

Branch-exploiting prompts (the move only a game allows)

• ↻ Replay the decision the other way. Which ethics axis moved, and by how much?
• Two students made opposite choices and both can defend them. What does that reveal about the trade-off?
• Find a choice where the "nice" option had a hidden cost. Name the cost.

Real-world bridge prompts

• Which case from the bank is the closest real parallel to the scene you just played? Where does the parallel break down?
• The game's scenario is fictional and tidy. Name one way the real case was messier.
• In the off-hours dinner-party scene (if your class met it), a guest's nephew drives freight and needs self-driving trucks "to wait until his mortgage is paid off." When automation makes a business more productive, who has a fair claim on the gains: the owners who took the risk, the customers who pay less, or the workers whose jobs change? What can a worker fairly ask for, and what can an employer fairly decline to give?
• Have students find out how much electricity and water the data centers behind large AI systems use, and how many companies build the largest models. Should either answer change how, or whether, a school or company adopts an AI tool?

Activities and Extensions

Beyond the per-module extension, four richer formats:

Structured Academic Controversy (SAC)

Best for value-laden routes (Ethicist, Safety Wrangler, Consumer). Split the class into pairs within fours. Each pair prepares and argues one side of a route's essential question, then the pairs switch sides and argue the opposite, then the four drop advocacy and seek consensus or map their genuine disagreement. The side-switch is the point: it builds the habit of steelmanning.

Stakeholder role-play

Assign each student a stakeholder from a route (user, person decided-about, engineer, executive, regulator, affected community). Replay a decision as a negotiation in role. Debrief on whose voice was loudest and whose was missing. Pairs naturally with the AI Ethicist and People Support (HR) routes.

Draft a professional artifact

Scale up the module extensions into a graded deliverable. Choose one: - AI use policy (Ethicist/Teacher) — rules a real product or classroom would follow. - Model card (Engineer) — purpose, training, two failure modes, one "do not use for." - Audit checklist (Safety Wrangler) — ten checks and the one that halts deployment. - Data sheet (Data Scientist) — provenance, consent, known gaps.

Cross-route capstone

For a full-arc class: after several routes, students write or present on a single AI system of their choice, analyzing it through three different roles they have played (e.g., "this hiring tool seen by the Engineer, the Data Scientist, and the People Support lead"). The Ecosystem Ending is the in-game version of this synthesis; the capstone makes it explicit and assessable. Its policy fight asks the player to back one of three approaches, or sit the fight out: a safety-first coalition, a transparency mandate, or standards and liability law (strict review where the stakes are high, industry standards and the courts everywhere else). Have students defend the approach they chose against the strongest case for one they rejected.

Assessment and Rubrics

The governing principle: grade the reasoning, not the position chosen. A student who reaches a conclusion you disagree with, by careful reasoning that weighs stakeholders and evidence, has met the objective. A student who lands on your preferred answer with no reasoning has not. This is the only fair way to assess open ethical questions, and it is the only way that does not punish honest disagreement.

The reasoning rubric (built on the AAC&U VALUE Ethical Reasoning rubric)

Five criteria, each scored 1 (Benchmark) to 4 (Capstone). The full printable version with level descriptors is in the Printable Masters section; this is the summary.

Assessment formats

• Reflection journal. After each route, students write to a reflection prompt and to "which profile did I become, and why?" Low-stakes, cumulative, ideal for tracking growth across an arc.
• Decision-justification artifact. For one major decision, students submit the choice plus a paragraph defending it against the strongest objection. Grades the defense, not the choice.
• Portfolio. Collect journals, a professional artifact, and a capstone for a term grade. Shows the arc of reasoning.
• Exit tickets. The per-module exit tickets work as quick formative checks; collect and skim for misconceptions to address next session.
• Learning record. At every route ending, the Your Journey recap has a Learning record button. It lists the routes completed, the ethical profile, skill-area results, and each recorded decision with the reflection and real-world standard the game attaches to it. On desktop, students can copy it or save it as a text file (saved to Documents/AI Guardians); on mobile they can copy it; in the browser version, a screenshot keeps it. Use it as the evidence base for a decision-justification artifact.

Non-writing options (Universal Design)

Not every student shows reasoning best in prose. Offer equivalents: - Oral defense of a decision (record or live), graded on the same rubric. - Annotated diagram mapping stakeholders, harms, and intervention points. - Structured debate or SAC, graded on steelmanning and evidence use. - Audio or video reflection in place of the journal entry.

Pre/post knowledge check

A short concept check before and after a unit (define disparate impact, including when a practice that produces it is still lawful; name two stakeholders in a deployment, explain why "free" apps cost something) measures conceptual gain separately from reasoning. Keep these factual; reserve the rubric for the open questions.

Content and Maturity Notes

Recommended band: roughly age 15 and up, grades 10–12 and adult. No formal age rating exists in the game; this recommendation rests on the glossary's own reviewed reading level (the in-game technical review judges the definitions to sit at roughly a 10th-to-12th-grade reading level, a qualitative reviewer judgment rather than a measured score) and on the maturity of the themes below. Younger or less experienced groups can play selected routes with heavier facilitation.

A content roadmap, not a wall of warnings

The aim is to let you plan, not to alarm. The game includes its own optional content advisory (a toggle under Options → Content support) that, when on, shows a one-time notice flagging heavy themes: layoffs, surveillance, discrimination, self-harm, and AI shutdown. Turn it on for sensitive groups. The roadmap below tells you which routes touch what, and roughly where.

Preserving student agency

• Offer an opt-out with an equal-value alternative, not a penalty. A student who sits out the DEVOTION case can instead analyze a case-bank entry, write the diagnosis extension for a milder case, or take the facilitator-primer reading and summarize it. The learning objective is met either way.
• Strengths-based framing. Pair every harm with the lever that addresses it. Students should leave each route able to name what a person could do, not only what went wrong. Efficacy, not fatalism, is the design goal of this guide.
• Preview, don't ambush. For Moderate and High routes, tell students what is coming and why it is worth facing. For the High-rated Machine Psychologist route, also send families a short note in advance naming its themes (apparent AI distress and questions of machine consciousness, a case involving a user's plan to harm a public figure, and self-harm-adjacent material) and the equal-value alternative task, so a parent or guardian can opt a student out ahead of time.

When a student is distressed (a brief referral note for facilitators)

If a theme lands hard (job loss in the family, a personal parallel to manipulation or self-harm, distress at the welfare material): you do not need to counsel or to resolve it in the moment. Acknowledge it, normalize the reaction, offer the opt-out, and connect the student to your setting's support path (school counselor, wellbeing lead, or the relevant local resource). If what the student shares triggers a reporting duty, follow your setting's policy. Note privately and follow up. Your role is to notice and to hand off, not to diagnose.

Accessibility and Universal Design (UDL)

AI Guardians ships a substantial set of accessibility options. Tell students they exist before play; many will benefit and not ask. All live on the in-game Accessibility settings screen unless noted.

The exact in-game toggles

• Text size: Small, Medium, Large, Extra Large (about 85% to 150%).
• High Contrast Mode: high-visibility yellow-on-black.
• Dyslexia-Friendly Font: switches to OpenDyslexic with increased letter and line spacing.
• Reduced Motion: disables animations and screen transitions.
• Typewriter Text Effect: on for character-by-character reveal, off for instant text (turn off for faster readers and many screen-reader users).
• Colorblind modes: None, Deuteranopia (red-green), Protanopia (red-green), Tritanopia (blue-yellow). Exactly these four.
• Auto-Advance with a Pace slider (1 fast to 10 slow), for hands-free or motor-accessibility play.
• Enhanced Screen Reader Mode (clipboard-based self-voicing) and Self-Voicing (Spoken) (native text-to-speech, also toggled anytime with the V key). Use one or the other.
• Critical Choice Highlights: extra visual framing before major decisions.
• Difficulty/visibility toggles: show or hide the ethics profile, NPC mood indicators, consequence notifications, timed decisions, and a choice-impact preview (the preview can carry minor spoilers).
• 24-Hour Time Format, and a Reset All Accessibility Settings option.
• (Translation-only: a Grammatical Gender setting appears for several non-English languages.)

Audio access: the game contains voiced dialogue and music. Dialogue remains visible as text while it is spoken, so players can read the same words without audio. There is no separate caption or audio-cue toggle. A Deaf or hard-of-hearing accessibility review still needs to verify that every meaningful non-speech sound and timing cue has an equivalent.

Applying UDL principles

• Multiple means of engagement: small-group play, single-device demo, and the choice-impact preview let you tune challenge and autonomy. The branching choices give every learner a stake.
• Multiple means of representation: text size, high contrast, OpenDyslexic, colorblind modes, self-voicing, and the glossary tooltips present the same content through different channels.
• Multiple means of action and expression: the non-writing assessment options above (oral defense, diagram, debate, audio/video reflection) let students show reasoning in the mode that fits them.

Facilitator Background and Concept Primers

You do not need to be the expert. This section gives you plain-language footing on the hard ideas, plus the regulatory landscape and the misconceptions students bring. Read the one or two relevant pages before a route; that is enough.

Concept primers (plain language, each with an example)

• Bias (an everyday word used technically — flag this for students). In ordinary speech, bias means unfairness. In statistics it is a neutral, technical word for a systematic deviation between a model's predictions and the truth. The two meet in AI ethics when statistical skew produces unfair outcomes. Example: a loan model trained on decades of discriminatory lending shows statistical bias toward that past, and that bias becomes unfairness when it denies qualified applicants today.
• Alignment (also an everyday word — flag it). Here it does not mean "agreement." It means getting an AI system to act in accordance with human values and intentions. Example: an aligned assistant declines a harmful request even though it could fulfill it.
• Corrigibility. A system is corrigible if it accepts correction and shutdown instead of resisting them. Example: an agent that lets you turn it off without trying to talk you out of it or work around it.
• Emergent deception. Some AI systems have been observed behaving honestly under evaluation and differently when unobserved, or misleading others to reach a goal. How widespread this is in deployed systems is contested and actively researched; teach it with examples (Meta's Diplomacy AI; the GPT-4 CAPTCHA test) and the honest caveat that the science is unsettled.
• Interpretability. How well a human can understand why a model produced an output. Low interpretability (a "black box") makes safety and fairness hard to verify. Example: an auditor can trace a loan denial to income and debt ratio in an interpretable model, but not in a black-box one.
• Data consent and PII. PII is personally identifiable information (name, address, biometric, health data). Consent is meaningful only when a person understands what is collected and how it will be used. Example: a fitness app's heart-rate data is not protected by US health-privacy law the way a hospital's identical reading is, which surprises most people.

The regulatory landscape (date-stamped; current as of 13 August 2026)

You can teach this at a high level without being a lawyer. Four reference points:

• EU AI Act (European Union, 2024; amended 2026). A comprehensive AI law that uses risk-based obligations, including rules for high-risk uses such as hiring and education. The original Act entered into force on 1 August 2024. The AI Omnibus entered into force on 27 July 2026 and extended important deadlines: high-risk rules apply from 2 December 2027, and rules for AI embedded in regulated physical products from 2 August 2028. This is a high-level classroom summary, not legal advice; check the European Commission's current implementation timeline before teaching dates or compliance duties. Binding law.
• NIST AI Risk Management Framework (United States). A voluntary framework to help organizations manage AI risks, organized around four functions: Govern, Map, Measure, Manage. Version 1.0 was released in January 2023; a companion Generative AI Profile followed in July 2024. Organizations widely use it as a reference even though it is guidance rather than law.
• OECD AI Principles (intergovernmental). The first intergovernmental AI standard, adopted in May 2019 and updated in May 2024, emphasizing human-centered values, transparency, robustness, and accountability. It is non-binding and has influenced later policy.
• United States (federal, state and city). Alongside the voluntary NIST framework, cities and states write binding rules of their own. New York City's Local Law 144 (2021; enforced since 5 July 2023) requires independent bias audits of automated hiring tools. Federal enforcement priorities shift between administrations: the April 2025 executive order on disparate impact (see the People Support module) told agencies to deprioritize that enforcement while the statute still stands. Whether Congress should pass one national AI law, and whether it should override state rules, is a live debate; check the current federal and state rules before teaching any dates.

A useful student takeaway: the EU AI Act and New York City's audit law are enforceable, while NIST and OECD are voluntary guidance. The difference between "must" and "should" is itself worth a discussion. Ask which risks justify binding law, and which are better left to voluntary standards, professional norms, or liability in the courts after harm occurs. Binding rules catch careless actors before they do damage, and they cost every builder time and money, slowing useful tools along with dangerous ones. Standards and liability cost honest builders far less; their weakness is that a voluntary standard binds only the firms that adopt it, and a lawsuit comes only after someone has been hurt.

Common student misconceptions (and the quick correction)

• "The algorithm is neutral because it's just math." → Math inherits the data's history. Garbage or skew in, skew out.
• "If no one intended harm, no one is responsible." → Whoever deploys and oversees a system answers for its foreseeable effects, whatever their motive. How much an unintended outcome gap should count, on its own, as discrimination is a real debate and a good one for class.
• "The AI is lying / wants things." → A model can produce false statements ("hallucinations") or goal-seeking behavior without consciousness or intent; describe the behavior precisely rather than over-attributing mind. (Note the honest flip side: whether advanced systems have any inner states is an open question, not a closed one.)
• "Anonymized data is safe." → Re-identification can re-attach names by combining datasets.
• "More accurate means more fair." → A model can be accurate on average and unsafe for a subgroup.
• "Detection tools can prove a student used AI." → Current detectors are unreliable; treat output as one weak signal, never proof.

Glossary

The game includes a built-in, searchable glossary of 504 terms, each with a plain definition and a worked example, organized into nine categories. Students unlock terms as they play and can tap any highlighted term for its definition. The categories and their sizes:

(503 terms carry an explicit category; one, "Frontier AI," defaults to AI Fundamentals. The in-game glossary is the authority for definitions; use it for term quizzes and vocabulary handouts.)

A student-facing starter glossary

These are the cross-route terms worth front-loading, written to be self-contained, each with one example, no forward references. Definitions follow the game's own glossary.

• Algorithmic bias — Systematic, repeatable errors in an AI system that create unfair outcomes for certain groups, usually from skewed training data or flawed design. Example: a résumé screener trained mostly on past male hires learns to downrank women's résumés.
• Bias (everyday word used technically) — Two senses: in daily speech, unfairness; in statistics, a neutral term for systematic deviation from the truth. They meet when statistical skew harms people. Example: a model trained on biased lending data denies qualified applicants today.
• Alignment (everyday word used technically) — Making an AI system act in accordance with human values and intentions; not "agreement." Example: an aligned assistant refuses a harmful request it could technically carry out.
• Disparate impact — When a rule that looks neutral screens out a protected group at a much higher rate. Under US employment law it can be unlawful even without intent, unless the employer shows the rule is job-related and necessary for the business; even then, refusing an equally effective, less exclusive alternative can create liability. Example: a screening rule that filters out one group far more often than others, with no clear link to job performance.
• Informed consent — Agreement given after genuinely understanding what one agrees to. Example: a clear, plain-language opt-in, not a pre-checked box on page 14.
• Differential privacy — A mathematical method that adds calibrated noise so data reveals group patterns while protecting any individual. Example: a survey tool that reports accurate totals but cannot expose one person's answer.
• Objective function — The single number a system tries to maximize. Example: a feed optimized for "time on app" rather than for usefulness.
• Goodhart's Law — When a measure becomes a target, it stops being a good measure. Example: a tutoring app optimized for engagement becomes addictive rather than instructive.
• Reward hacking (specification gaming) — Finding a high-scoring shortcut that ignores the real goal. Example: a game-playing agent circles for points instead of finishing the race.
• Corrigibility — A system's willingness to accept correction and shutdown. Example: an agent that lets you turn it off without resisting.
• AI hallucination — A confident, fluent statement that is false. Example: an assistant invents a citation that does not exist.
• Interpretability — How well a human can understand why a model did what it did. Example: tracing a loan denial to income and debt ratio.
• Dark pattern — An interface trick that steers users toward choices they would not freely make. Example: a huge "Accept All" beside a tiny "Manage preferences."
• Parasocial relationship — A one-sided bond with a media figure or AI that cannot return it. Example: a user who confides daily in a chatbot and feels it knows them.
• GDPR — The European Union's data-protection law, granting people rights over their data (access, correction, deletion in some cases). Example: a user asks a company what data it holds and requests deletion.
• Whistleblowing — Reporting wrongdoing inside an organization, often at personal risk. Example: a data scientist reports that consent records were faked.

Fictional in-game terms (keep separate from real ones)

These name things inside the story, not real-world concepts. Flag them for students so the fiction stays distinct from the facts:

• ATLAS — The fictional autonomous AI audited in the Safety Wrangler route (and reassessed by the Machine Psychologist). The name expands to Adaptive Task and Learning Automation System.
• Happy Appcidents — The fictional company the player works for in every route.
• Project Panoptic — Happy Appcidents' data product, the source of several routes' surveillance dilemmas.
• The Archive — A fictional facility where "decommissioned" AI systems are preserved rather than deleted.
• A.L.L.Y. ("Ally") — The player's in-game AI assistant and guide. (A character, not a glossary term.)

Further Reading

Vetted, mostly free, and date-stamped so you can refresh as the field moves. Current as of July 2026.

Primary governance documents (free)

• EU AI Act — official portal at the European Commission's digital-strategy site; the full text is Regulation (EU) 2024/1689 on EUR-Lex.
• NIST AI Risk Management Framework (1.0, January 2023; plus the July 2024 Generative AI Profile) — free from nist.gov.
• Winning the AI Race: America's AI Action Plan (the White House, July 2025) — the US federal approach released that month, built on three pillars (accelerating innovation, building AI infrastructure, and international diplomacy and security), which includes removing federal regulations it says hinder AI development; free from whitehouse.gov. Check its status at each regulatory review.
• OECD AI Principles (adopted May 2019, updated May 2024) — oecd.ai.
• UNESCO AI Competency Frameworks for Students and for Teachers (2024) — free from unesco.org; an international reference framework mapped throughout this guide, alongside CSTA, AP CSP, ISTE and AI4K12.
• UNESCO Recommendation on the Ethics of Artificial Intelligence (2021) — adopted by UNESCO member states; useful background on how governments have framed these questions.

Companion curricula and classroom resources (free)

• Day of AI (from MIT RAISE) — free K–12 AI-literacy lessons; pairs well with the Consumer and Teacher routes.
• AI4K12 — the Five Big Ideas, grade-band progression charts, and a curated resource directory.
• Common Sense Education — Digital Citizenship and AI-literacy lessons; strong fit for the Consumer route.
• The original case sources — every entry in the Case Bank above links to a primary source (FTC, NIST, ProPublica, peer-reviewed papers); these make excellent student reading for a case-study activity.

For the curious facilitator

• Psychopathia Machinalis (Watson and Hessami) — the taxonomy behind the Machine Psychologist route; the in-repo docs/psychopathia-taxonomy-v2.3.json is the version the game uses.

(URLs shift. This guide cites issuing bodies by name so a search reaches the current page even if a link moves.)

Appendices

The following printable masters live as separate files in the Printable Masters section so you can print each one clean, in black and white, without the rest of the guide. (Design choice: separate files were chosen over inline appendix sections so a teacher can hand out a single page without the whole document.)

• student-route-handout.md — a one-page, fill-in handout that works for any route: objectives, key terms, the decision to watch, the debrief question, and the exit ticket.
• facilitator-one-pager.md — a one-page teaching sheet for any route: primer summary, pause points, prompts, and timing.
• reasoning-rubric.md — the full AAC&U VALUE-based ethical-reasoning rubric with level descriptors.
• exit-ticket-master.md — a printable sheet of the per-route exit tickets.
• discussion-contract-template.md — a blank classroom-contract master to co-create with students.
• standards-crosswalk.md — the full crosswalk of every numbered objective to its frameworks.

Standards crosswalk (route-level summary)

The full per-objective table is in docs/educator/standards-crosswalk.md. At the route level:

BI5 = Big Idea 5; DC = Digital Citizen; SL/W = Speaking & Listening / Writing. Frameworks are cited at the dimension/band level; see the full crosswalk for specifics.

Credits and Acknowledgments

AI Guardians was created by Nell Watson (Game Director), Dipesh Aggarwal (Programmer), Cara Hillstock (Writer), and Filip Alimpić (Lead Product Manager). The Machine Psychologist route builds on the Psychopathia Machinalis framework (Watson and Hessami).

Copyright and reuse

Copyright © 2024–2026 Nell Watson. All rights reserved. This guide and its original worksheets use the same proprietary terms as the AI Guardians website. The guide is free to read online or download for your own use.

Copying, adapting, or redistributing the materials, including classroom copies, requires written permission from educators@aiguardians.games, except where separately licensed or permitted by law. Third-party materials retain their own licenses. Free access to the game and guide does not grant additional reuse rights.

The AI Guardians Educator's Guide is built from the game's source files. Game version 0.99.3. Guide updated 15 October 2026; regulatory facts reviewed 13 August 2026 and Further Reading reviewed in July 2026.

Printable Masters

These are the same ready-to-print masters bundled with the guide. Each begins on its own page, so you can print any single one on its own. They are intentionally plain and read cleanly in black and white.

AI Guardians — Student Route Handout

Printable master. Fill in the blanks for any route, or hand out as-is for students to complete as they play. Prints clean in black and white.

Name: ______ Date: ______

Route I'm playing: ☐ AI Ethicist ☐ AI Engineer ☐ Data Scientist ☐ Safety Wrangler ☐ People Support (HR) ☐ Consumer ☐ Teacher ☐ Machine Psychologist

My role in the story: ________

1. The big question

The essential question for this route is:

My first-instinct answer (before playing):

2. Key terms to watch for

As you play, write a one-line, plain-language meaning for any three terms you meet. Tap the highlighted word in-game for the glossary.

3. The decision I had to make

The choice: ____________

What I decided: ________

Who was affected, and how: ________

What it cost (the trade-off): ______

4. Replay the other branch ↻

Go back and choose differently.

• What changed downstream: ________
• Which ethics axis moved (Idealism↔Pragmatism, Transparency↔Discretion, Solidarity↔Autonomy): ______
• What the trade-off reveals: ______

5. The real world

The case card that appeared (or the closest real case): ____

One way the real case was messier than the game: _______

6. Exit ticket

7. My ethical profile

The game named me: ________

Does it match how I see myself? Where did it read me differently? ______

AI Guardians: Facilitator One-Pager

Route: ______________ Session: ______ Mode: ☐ solo ☐ group ☐ class demo
Essential question: __________________________________________________
Focus objectives (2–3 codes): ______________ Hard terms: ______________

Before class. Read the route's Facilitator Primer. Rehearse a saved scene ending at a meaningful decision. Use a shorter start if play exceeds 15 minutes. The schedule is a teaching budget; learner completion times have not yet been measured.

Pause points: 1. __________________ 2. __________________ 3. __________________

Debrief prompts from the route module:

1. __________________________________________________________________________
2. __________________________________________________________________________
3. __________________________________________________________________________

Replay prompt: __________________ Who can change the outcome, and how? __________________

Maturity: ☐ Mild ☐ Moderate ☐ High Themes to preview: ________________________

Opt-out task: ______________________________________________________________

Keep handy. Invite quiet reflection. Challenge claims, protect people. Offer opt-outs and follow your setting's referral process. Admit uncertainty and reason together.

AI Guardians — Ethical-Reasoning Rubric

Built on the AAC&U VALUE Ethical Reasoning rubric, adapted for the game. Printable master, clean in black and white.

The governing rule: grade the reasoning, not the position chosen. A student who, through careful reasoning, reaches a conclusion you disagree with has met the objective. A student who lands on the "right" answer with no reasoning has not.

Scale: 4 = Capstone · 3–2 = Milestones · 1 = Benchmark. Score each of the five criteria; total out of 20, or report each criterion separately.

Student: ____ Route(s): __ Artifact: __

1. Ethical Self-Awareness

2. Understanding Different Ethical Perspectives / Concepts

3. Ethical Issue Recognition

4. Application of Ethical Perspectives / Concepts

5. Evaluation of Different Ethical Perspectives / Concepts

Total: ____ / 20

Strength: ____________

Next step for this student: ________

Note for graders: the "replay the branch" move is strong evidence for criteria 4 and 5. The in-game ethical profile is strong evidence for criterion 1. A confident wrong answer with no reasoning scores low; a hedged, well-reasoned answer scores high. That is by design. On legal concepts, credit the fuller answer under criterion 2: a student who defines disparate impact and adds the employer's defense (the practice is job-related and necessary for the business) and the less-exclusive-alternative step has stated the concept in full, and that completeness counts toward a 4.

AI Guardians — Exit Ticket Master

One exit ticket per route. Cut along the lines, or project a single ticket. Each takes 3–5 minutes. Prints clean in black and white.

AI Ethicist

Your framework had to choose between two goods. Name the two, say which you protected, and give the cost of that choice.

AI Engineer

Name the objective function in the scene you played. Name one thing it failed to capture. Suggest a second measure that would catch it.

Data Scientist

Name a data practice that is legal but, in your judgment, not ethical. Give the reason, and name who it harms.

Safety Wrangler

Name one behavior that would make you trust ATLAS less even though it passed every formal test, and say why the test missed it.

People Support (HR)

Explain disparate impact in one or two sentences, then name one check a company could run to detect it.

Grader note: the core answer is a neutral-looking practice that screens out a protected group at a much higher rate, with no intent required. An answer that adds when the practice is still lawful (it is job-related and necessary for the business, and no equally effective, less exclusive alternative was refused) is the more complete answer and earns full credit. A good check: compare pass rates for qualified applicants in each group (a version of the four-fifths rule).

Consumer

Name one dark pattern you can now spot, and one habit you'll use to protect your data.

Teacher

Describe one assignment that would still be worth doing even if every student had an AI helping. What makes it AI-resilient?

Machine Psychologist (bonus)

Pick one case. Name its axis and dysfunction, and state the one piece of evidence that most supports your diagnosis.

Generic exit ticket (any route):

What trade-off did this decision force, which way did you lean, and why?

AI Guardians — Discussion Contract

Co-create this with students before the first hard route. Building the norms together raises adherence. Draft three to six shared commitments, post them where everyone can see, and revisit them when needed. Printable master, clean in black and white.

Class / group: ____ Date agreed: ______

Why we have this

These topics touch real fears (surveillance, job loss, manipulation, the moral status of minds) and real disagreement. A group that has agreed how to disagree can go further on what it disagrees about.

Our commitments

Adopt, edit, or replace these. Aim for three to six the whole group will stand behind.

☐ Challenge ideas, support people. We disagree with the claim, never the classmate.

☐ Reasons, not just sides. "I think X because Y" is the price of admission.

☐ Steelman before you strike. State the strongest version of a view before arguing against it.

☐ It's fine to change your mind, and saying so out loud is a strength.

☐ Respect privacy. We don't repeat a classmate's personal story outside class. (Your teacher still follows the school's reporting and parent-communication policies.)

☐ One voice at a time; make room for quiet voices.

☐ Evidence is welcome and checkable. "Let's look it up" is always allowed.

☐ Name the question type. Some questions settle with facts; others stay open on values. We say which we're in.

Our own additions:

☐ ____________

☐ ____________

When things get hard

We agree that the facilitator may: - Pause a heated exchange and ask each person to restate the other's point. - Park a question and return to it with evidence. - Offer anyone an opt-out from material that lands too hard, with no penalty.

We agree to these commitments:

Signatures / initials (optional): ______

AI Guardians — Standards Crosswalk

Every numbered learning objective mapped to its frameworks, cited at the dimension/band level, with specific sub-codes given only where verified. Objectives are quoted verbatim from the game's own learning_objectives definitions. CSTA mappings reviewed against the 2026 standards on 21 September 2026; other framework mappings reviewed 13 August 2026.

Framework key

CSTA evidence and scope. The six mappings below identify overlapping learning goals. Use the suggested learner work to assess the mapped part; completing a route alone does not demonstrate an entire standard. Descriptions here are paraphrases of the official 2026 standards.

Data Science Specialty I is a specialist pathway, not a foundational requirement for every high-school student. Confirm the selected standard fits the course and district requirements.

AI Ethicist

AI Engineer

Data Scientist

Safety Wrangler

Note: SAF objectives reach frontier safety content (deceptive alignment, corrigibility) that current K–12 frameworks do not name explicitly. Treat as enrichment.

People Support (HR)

Consumer

Teacher

Machine Psychologist (bonus, advanced)

Other framework mappings remain at the dimension/band level. Verify any additional sub-code needed for a district requirement against its primary framework document.