TFIS — AI Ethics in Education 3326 words

TFIS — AI Ethics in Education

Course Syllabus · v1.0
Full title Course: Navigating Ethical Challenges of AI in Student Learning
Audience Higher-education / pre-university educators, academic staff, programme leaders (Malaysian HE)
Format 7 live sessions × 2 hours = 14 contact hours, 1–3 Sep 2026 (Tuesday–Thursday) + Session 8 action-plan presentations (11.00 am–1.00 noon, Day 3)
Primary context sources/course-ai-ethics-course-pack.md (Facilitator Course Pack v1.0) — run sheets, scripts, prompt library P01–P15, instruments D1–D7
Sources Restricted register: Stanford HAI (2026 AI Index) primary; Anthropic, Penn/Wharton, MIT, Oxford, Harvard secondary. See sources/source-register.md. No other sources cited anywhere in the course.
This document The contract. Every deck, activity, facilitator guide, and rubric must satisfy what is specified here. Deviations get patched here first.

1. The capstone (stated first, because everything serves it)

Every participant leaves with a draft institutional AI-use guideline (v0.1) for their department — two pages maximum, seven components, version number, named owner, review date, red-team backlog, and its own AI-use disclosure statement.

2. Course-level learning outcomes

By the end of the course, participants can:

3. Design commitments (binding on all downstream deliverables)

  1. Case-first for S2 and S3. Both sessions open on a real incident before any framework is named. Principles and evidence are extracted from the case discussion, never lectured ahead of it. Specific openings are fixed in the session blocks below. No slide may present the five principles (S2) or the detector verdict (S3) before the room has worked the opening case.
  2. The pack is the spine. Session structure, prompts (P01–P15), instruments (D1–D7), and scripts come from the course pack. The syllabus adjusts ordering only where the case-first commitment requires it (S2, S3).
  3. Two spines run throughout. Competency: the 4Ds (AI Fluency Framework, CC BY-NC-SA — legally remixable, and we disclose that we do). Maturity: the EAI-CMM, administered S1 and S7; institutional variant in S6.
  4. S6–S7 policy grounding is Malaysian. Reuse the AI Governance Bill fact-sheet work from the adjacent TFIS project: PDPA 2010 as the statutory data floor, risk-tier framing for the permission architecture, national AI-governance direction for the exemplar scan. [LOCAL] slots mark where MQA/MOHE/senate specifics get filled by the guideline's named owner.
  5. Slide format. All decks are single-file HTML on the tfis-fluency brand system (cream/pine/sun palette, Archivo + Hanken Grotesk, existing slide/player pattern). One file per session, 02-slides/s1.html … s7.html.
  6. Energy rule. S2 and S6 are 8–10 pm slots: discussion-heavy, no lecture block over 12 minutes. Decks for these two sessions carry fewer content slides and more work-surface slides.
  7. Language. Delivery in English or Bahasa Melayu as the room prefers; artifacts drafted in the language the institution's policies actually use.
  8. Live-demo fallback. Every live AI demo (P01, P03, Learning Mode) ships with offline screenshot fallbacks in the deck appendix.

4. Session-by-session specification

Session 1 — The AI Revolution in Higher Education: Opportunities and Ethical Challenges

Tuesday 1 Sep · 2.30–4.30 afternoon

Objectives. Participants can (1) describe the AI capability frontier with 2026 AI Index evidence, (2) establish their personal baseline on the EAI-CMM, (3) map gifts and threats for their own discipline, (4) state axiom A1: the gap is the curriculum.

Key topics. Cold-open live demo (a real assignment from the room, completed by AI in ~90 seconds, P01) · capability acceleration vs the jagged frontier (IMO gold vs 50.1% on analog clocks) · adoption data (4 in 5 students using AI; only 6% of teachers call their policies clear) · what students actually do (Anthropic Education Report: ~half of use is direct answer-seeking) · the opportunity ledger (Tutor CoPilot, guardrailed tutoring, CS50, Learning Mode demo).

In-session activity. Threat/Gift Matrix (30 min): discipline groups fill a 2×2 (Teaching/Assessment × Gift/Threat), circle the most urgent threat and most undervalued gift; master chart stays on the wall all three days. Plus EAI-CMM baseline (15 min, private) and exit ticket 3-2-1.

Capstone contribution. Baseline EAI-CMM score (the delta S7 measures against) · Threat/Gift master chart = the raw risk/opportunity inventory the guideline's scope section must answer · fixed teams formed — these become the S6–7 drafting teams.

Readings / cases. 2026 AI Index Report, Education chapter (primary pre-read) · Anthropic Education Report (Apr 2025) · Wang & Demszky, Tutor CoPilot (2024) · optional pre-work: AI Fluency Framework & Foundations course.

Session 2 — Understanding AI Ethics in Student Learning: Principles, Risks and Responsibilities

Tuesday 1 Sep · 8.00–10.00 pm (evening rule applies)

CASE-FIRST OPENING (binding). No principles slide first. Open on a real biased-selection incident, worked cold: the UK Ofqual 2020 grading algorithm (39% of teacher-assessed grades downgraded, disadvantaged schools hit hardest, national U-turn) with the St George's Hospital Medical School admissions program (1980s automated screening that penalised women and non-European names — the original biased-admissions model) as the historical anchor: same failure, forty years apart. Room task before any framework: what exactly went wrong, who was harmed, who was responsible? The five principles are then named as labels for what the room already said.

Objectives. Participants can (1) extract ethical principles from a concrete incident rather than recite them, (2) apply the five-principle lens (Floridi et al.: beneficence, non-maleficence, autonomy, justice, explicability) to AI-in-learning scenarios, (3) demonstrate bias empirically rather than rhetorically, (4) assign responsibility across the educator–student–institution triad.

Key topics. The opening case pair → five principles extracted · live bias demo (P03: Ahmad/Aisyah reference letters; adjective delta; audit-per-model habit; Bahasa Melayu / kampung-vs-suburb framing) · the six-risk map, one line of evidence each (integrity, learning/crutch preview, cognitive [MIT preprint, cited with limitations], equity, privacy/PDPA, wellbeing/dependence) · the responsibility triad.

In-session activity. Ethics Triage (30 min): eight scenario cards ranked on a severity ladder, tagged by primary principle and primary owner; strict ranking, no ties; inter-group argument on divergences. Card 4 (detector-only failing grade) deliberately splits the room to set up S3. Case memo (≤200 words) assigned, due S3.

Capstone contribution. The five principles become guideline component 2 (Principles, localized) · triage severity logic seeds the integrity procedure (component 5) · the responsibility triad prefigures ownership (component 7).

Readings / cases. Floridi et al., AI4People (2018) · Wan et al., EMNLP Findings 2023 (reference-letter bias) · Kosmyna et al., MIT (2025 preprint — teach its limitations) · Bastani et al. PNAS abstract (preview) · Stanford HAI "delusional spirals" coverage (2026) · case dossiers: Ofqual 2020, St George's (to be compiled in sources/cases/).

Session 3 — Academic Integrity in the AI Era: Managing Plagiarism, Bias and Copyright Issues

Wednesday 2 Sep · 8.30–10.30 am

CASE-FIRST OPENING (binding). Open on a real false-accusation incident: the Texas A&M–Commerce May 2023 case (instructor ran essays through ChatGPT asking "did you write this?", moved to fail an entire section on that basis; diplomas held) — worked cold: what was the evidence, would it survive your institution's appeals process, who bears the burden? Only after the room has convicted the method does the Stanford Patterns data land: 61%+ of real TOEFL essays falsely flagged; detectors near-perfect only on native-speaker text; one laundering prompt flips the verdict. In a Malaysian university, second-language writers are not the edge case — they are the main case. Case memos from S2 are read before the opening (continuity), exit-ticket questions answered.

Objectives. Participants can (1) explain why detector output fails as evidence, citing the false-positive bias mechanism (perplexity ≈ lexical variety ≈ punishing L2 writers), (2) reframe integrity from artifact-policing to process-evidence design, (3) run a fair, non-accusatory suspected-misuse conversation, (4) map the current copyright fault lines (training data, output ownership, student IP + PDPA).

Key topics. What plagiarism rules were a proxy for · the detector evidence (bias + evadability = worst possible evidentiary instrument; researchers' own recommendation: don't use in evaluative settings) · misuse is still real (Anthropic Education Report: answer-seeking and detector-evasion requests at scale — no swinging to denial) · the successor regime: version trails, viva sampling (P06), in-class anchors, disclosure-as-norm (P05, drafted live) · copyright fault lines · the scripted misuse conversation ("walk me through how you made this").

In-session activity. Detector on Trial (30 min): structured moot on the motion "detector scores are admissible primary evidence," with sides deliberately assigned against participants' voiced views (steelmanning). Judges rule with one-sentence ratio. Expected convergence: at most a screening signal, never proof. Redesign warm-up ticket: name the assessment you'll rebuild in S5.

Capstone contribution. Disclosure template (P05 output) → guideline component 4 verbatim · detector-demotion position with the Patterns citation → component 5 (policies with footnotes get challenged less) · copyright/PDPA questions → component 6 · warm-up ticket = S5's raw material.

Readings / cases. Liang et al., Patterns 2023 (the evidentiary core) · Stanford HAI news write-up of the same · Anthropic Education Report (misuse patterns) · Harvard CS50 (Malan et al.) as the design-not-policing exemplar · case dossier: Texas A&M–Commerce 2023 · copyright: principles-not-cases note (litigation unstable).

Session 4 — Hands-on Workshop: Ethical AI Tools for Teaching, Assessment and Student Learning (Part 1)

Wednesday 2 Sep · 11.00 am–1.00 noon

Objectives. Participants can (1) apply the 4D framework to a real teaching workflow, (2) run the Description↔Discernment loop through ≥3 logged iterations, (3) produce two working, verified prompt workflows (rubric builder, feedback assistant), (4) keep a verification log as a professional artifact.

Key topics. The 4Ds in twelve minutes (delegate production, never judgment) · facilitator models the loop live (P07 thin-vs-full prompt, discernment narrated aloud) · target ratio 20 min instruction / 100 min doing.

In-session activity. Lab 1 — Rubric builder (35 min, P07, real assessment from a course taught this semester, ≥3 verification-log rows). Lab 2 — Feedback assistant (35 min, P08, past anonymized excerpt only — PDPA practice as lab rule; AI drafts, educator authors; no grades). Gallery: three volunteers project their logs, not their outputs.

Capstone contribution. Verification-log habit → the guideline's staff-diligence norm (component 4 symmetry: staff disclose too) · working workflows prove what "staff AI use" clauses must permit rather than prohibit · anonymization rule rehearses component 6 before it's written.

Readings / cases. AI Fluency Framework (Dakan, Feller & Anthropic) · Anthropic faculty-usage report (curriculum design as top faculty use case) · pack Appendix A (P07–P08) and D2.

Session 5 — Hands-on Workshop: Ethical AI Tools for Teaching, Assessment and Student Learning (Part 2)

Wednesday 2 Sep · 2.30–4.30 afternoon (keystone session)

Objectives. Participants can (1) explain the crutch effect from the RCT evidence, (2) redesign an AI-vulnerable assessment with the three-lane pattern, (3) build and break-test a guardrailed Socratic tutor, (4) write the student-facing AI-policy paragraph for one real course.

Key topics. The one study to remember — Bastani et al., PNAS 2025: +48%/+127% assisted, then GPT-Base students −17% vs never-had-AI unassisted; guardrails engineered the harm away; performance ≠ learning; design determines outcome; the dashboard lies in one direction · seven roles for AI in learning (Mollick & Mollick) · the three-lane pattern (🔴🟡🟢) as the course's core design vocabulary.

In-session activity. Lab 3 — Redesign sprint (40 min, pairs): audit your own assessment with P09, assign lanes, swap-test attack, patch. Lab 4 — Guardrailed tutor (30 min, P10): build, then social-engineer your own tutor three ways; passing = it holds. Lab 5 — Policy paragraph (15 min, P11): ≤150 words, first-year-actionable test. Exit ticket: one word — your redesign's weakest point (feeds S6 gap analysis).

Capstone contribution. Three-lane vocabulary → guideline component 3 (the permission architecture) · redesign sheet → assessment-design guidance the institutional scan checks for (S6 item 4) · policy paragraph = the course-level cell of the institutional policy; S7 explicitly promotes personal practice into institutional text · portfolio checkpoint: four artifacts in hand.

Readings / cases. Bastani et al., PNAS 122(26) 2025 (the citable RCT) · Mollick & Mollick, Assigning AI (2023) · pack D3, P09–P11 · cross-reference: MIT preprint vs PNAS RCT as a discernment lesson (which one survives senate?).

Session 6 — Developing Institutional Guidelines for Ethical AI Use in Teaching and Learning (Part 1): Diagnose

Wednesday 2 Sep · 8.00–10.00 pm (evening rule applies)

Objectives. Participants can (1) score their institution on the 10-item institutional maturity scan, (2) name the seven components of a complete AI guideline, (3) evidence their institution's three largest gaps, (4) enter S7 with an agreed drafting brief.

Key topics. From craft to constitution (heroism is what institutions run on when governance is absent; only 6% say policies are clear) · axiom A7: policy is a product · policy anatomy: the seven components (scope/definitions, principles, permission architecture, disclosure standard, integrity procedure, data & privacy rules, ownership & review) · Malaysian grounding (binding): reuse the AI Governance Bill fact-sheet work — PDPA 2010 as the statutory floor for component 6; risk-tier thinking mapped onto the three-lane permission architecture (higher-stakes use ⇒ tighter lane + more process evidence); national AI-governance direction and [LOCAL] MQA/MOHE slots · what good looks like: devolved-but-scaffolded pattern (Stanford/Harvard/MIT), default rules matter more than ideal rules (decide tonight what silence means), the teachable-policy test.

In-session activity. Institutional maturity scan (20 min, teams, 10 items × 0–4, bands harvested on the board — typical Malaysian HE 2026 result: L1–L2). Gap analysis (35 min, P12): three lowest areas × (current state · target component · cost of inaction as a concrete scenario from this course's evidence — committees move on scenarios, not scores).

Capstone contribution. Direct build input: scan scores + ranked gap analysis + chosen default rule + sketched disclosure format + nominated [LOCAL] owner = the drafting brief, S7's entry ticket. No brief, no draft.

Readings / cases. TFIS AI Governance Bill fact sheets (adjacent project — to be placed in sources/) · PDPA 2010 practitioner summary · Stanford/Harvard/MIT generative-AI teaching guidance pages (shape references) · pack §6 exemplar-scan talking points.

Session 7 — Developing Institutional Guidelines for Ethical AI Use in Teaching and Learning (Part 2): Ship

Thursday 3 Sep · 8.30–10.30 am

Objectives. Participants can (1) produce a complete seven-component Guideline v0.1, (2) conduct and survive a four-persona red-team review, (3) quantify their three-day capability delta on the EAI-CMM, (4) convert the guideline into a personal 7/30/90-day action plan with a kill criterion.

Key topics. Sprint rules (45 min, ≤2 pages, version block on page one; AI used heavily — this is a 🟢 task under Tuesday's discipline; the guideline carries its own AI-use disclosure) · hard constraints: three-lane vocabulary used, detector-evidence status explicit with citation, PDPA clause present, [LOCAL] slots owned · red-team personas: laundering student, overworked lecturer, falsely accused, auditor · v0.2 backlog as product honesty · EAI-CMM retake and delta (+8–18 expected, artifact-driven) · 90-day plan: guideline moves one institutional step with a named ally and a date; kill criterion mandatory.

In-session activity. Drafting sprint (45 min, P13) → red-team exchange (25 min, P14, written findings, severity-ranked, no oral debate) → patch round (15 min, top three findings fixed, rest logged) → EAI-CMM retake → action-plan skeleton (P15).

Capstone contribution. This is the capstone landing: Guideline v0.1 shipped, attacked, patched, versioned, owned — plus the measured delta and the 90-day deployment plan. Session 8 (Thursday 3 Sep · 11.00 am–1.00 noon) presents artifacts against the summative rubric (pack §S8: evidence discipline, design over policing, deployability, ethical reasoning, falsifiability; pass ≥12, distinction ≥17).

Readings / cases. Pack D5–D7, P13–P15 · S8 summative rubric · participants' own S1–S6 artifacts (the real reading list for this session).

5. Assessment map

Instrument When Type Feeds
EAI-CMM (20 items, 5 pillars) S1 baseline, S7 retake Self-assessment, private Delta = the course's measurable promise
Exit tickets (3-2-1; one-word) S1, S5 Formative S3 opening; S6 gap analysis
Case memo (≤200 words) S2→S3 Formative S3 opening; integrity-procedure thinking
Verification log (≥3 rows) S4 Formative, collected Staff-diligence norm in guideline
Redesign sheet + tutor break-test S5 Formative, portfolio Assessment-design guidance; component 3
Policy paragraph (≤150 words) S5 Formative, portfolio Course-level layer of the guideline
Drafting brief S6 Gate (entry ticket to S7) The build
Guideline v0.1 + backlog S7 Capstone Session 8 presentation; the institution
7/30/90 action plan + kill criterion S7 Summative Session 8; post-course follow-through
Summative rubric (5 criteria × 4 bands, max 20) Session 8 (Thursday 3 Sep) Summative Pass ≥12 · Distinction ≥17 · one "Ship It" award

6. Build status — all downstream deliverables verified complete

Deliverable Status Location
Decks 🟢 All 7 built 02-slides/s1.html … s7.html (2,405 lines combined) — cream/pine/sun, Archivo/Hanken, S2/S6 obey evening rule, S2/S3 order case-first, demo-fallback screenshots in appendices
Facilitator guide 🟢 Built 03-facilitator-guide/facilitator-guide.md — run sheets adapted from pack, S2/S3 case-first reordering, logistics checklist
Prompt handouts 🟢 Built 03-facilitator-guide/prompt-handouts.md — P01–P15, print-ready, model-agnostic
Assessment instruments 🟢 D1–D7 all built 04-assessments/templates.md (D1–D6 + S8 rubric + scenario cards) · 04-assessments/eai-cmm.md (D7: EAI-CMM 20-item + institutional variant) · 04-assessments/capstone-brief.md (standalone one-pager)
Course pack 🟢 v1.0 placed sources/course-ai-ethics-course-pack.md (739 lines) — full scripts, talking points, prompts, appendices
Source register 🟢 Placed sources/source-register.md — restricted set, per-session mapping
Case dossiers 🟢 All 3 compiled sources/cases/ofqual-2020.md · st-georges-1980s.md · texas-am-commerce-2023.md
AI Governance Bill grounding 🟢 Placed sources/AI_Governance_Bill_Fact_Sheet.md (NAIO 7 fact sheets) · sources/malaysia-ai-governance-note.md (crosswalk to course)
Print pack 🟢 Built 05-print-pack/print-pack.html + print-pack.pdf — A4 printable, cover page, all instruments

Syllabus v1.0 · 7 Jul 2026 · Updated for 1–3 Sep 2026 delivery · Owner: Khalil (TFIS) · Review: after first delivery, 4 Sep 2026 · Drafted with AI assistance under human direction; sources restricted to the register — practicing the diligence the course preaches.