Lesson Report:
Title: Trustworthy AI in Policy Practice: From Student Proposals to EU Principles and Real-World Dilemmas
Synopsis: The session connected student policy memos on AI governance to the EU’s “trustworthy AI� framework, focusing on how abstract principles translate into concrete design and regulatory choices. Through breakout analyses and applied case discussions (self-driving cars and post-attack facial recognition), the class examined trade-offs between autonomy, prevention of harm, fairness, and explicability, and previewed next steps to refine existing policy proposals.
Attendance
– Absent: 2 (Amin, Jonah)
Topics Covered (chronological)
– Lightning policy memo presentations (2 minutes each)
– Mohamed Omar: Opaque AI in California government decisions
– Problem: Black-box automated decisions in law enforcement and public benefits; GAO audits (2021, 2023) flag transparency gaps; unaccountable police facial recognition already in use.
– Proposal: Amend CCPA to grant residents a legally enforceable right to a meaningful explanation for any significant automated government decision; enforce via CA Attorney General.
– Rationale: Advances due process and accountability; leverages existing privacy framework; standard: if AI can’t explain, it shouldn’t decide.
– Objections/responses: Innovation chill rebutted—this is accountability, not a ban; markets can innovate to meet standards. Cost of inaction: wrongful denials, public trust erosion.
– Instructor note: Useful model for “amend an existing ruleâ€� activity later.
– Freshta: AI-generated disinformation in U.S. elections
– Problem: Deepfakes, fabricated articles, and AI political ads erode public trust and informed choice; rapid social media spread complicates verification.
– Proposal: FTC “AI content transparency and labelingâ€� rule; mandatory labeling of synthetic political content; public FTC database of AI-generated political materials for verification.
– Rationale: Transparency/accountability without restricting speech; addresses public concern (Pew: >50% Americans fear AI will influence elections); helps voters distinguish human vs machine-made.
– Samira: AI disinformation in EU elections
– Proposal: “EU Digital Election Integrity Ruleâ€� requiring labeling of AI-made/modified political content; platforms and campaigns must report AI use; penalties for noncompliance; create EU AI transparency task forces to oversee.
– Goal: Fairer, more transparent elections; anticipate pushback from large platforms but prioritize freedom, fairness, and democratic trust.
– Framing today’s objectives
– Connect recurring memo themes (especially disinformation/labeling) to formal policy frameworks.
– Reading anchor: EU guidelines for trustworthy AI.
– Agenda:
1) Unpack each EU principle in groups with examples.
2) Apply a single principle (prevention of harm) to a hard case (self-driving car).
3) Explore inter-principle conflict (post-attack facial recognition) and preview amending/tweaking existing rules.
– Activity 1: Unpacking the EU’s four principles (5-minute breakout, one per group)
– Respect for human autonomy (Group 1: Bekaim, Elijah, Gavin, Mohamed Omar)
– Core meaning: Humans should retain decision-making power; AI decisions should be subject to human oversight/override (“human-in-the-loopâ€�).
– Tensions: Capitalist efficiency/profit incentives can sideline human judgment; fear of a drift toward technocracy.
– Examples/concerns: Deepfakes; job displacement (e.g., automation at Amazon); “AI ministerâ€� anecdote raising legitimacy/authority questions.
– Instructor emphasis: Autonomy = preserve the ability to make and override decisions; human input prioritized.
– Prevention of harm (Group 2: Izirek, Amira, Safi, Wuti)
– Scope: Avoid causing/contributing to physical, psychological, social, economic, and environmental harm; protect vulnerable groups.
– Ideal: Safe, secure, resilient systems (e.g., self-driving cars that do not crash and adapt to conditions, protecting children, elderly, disabled).
– Instructor extension: This principle would exclude entire product classes designed for harm (e.g., military AI). Mentioned Anduril/Palantir as reference points.
– Fairness (Group 3: Freshta, Samira, Banu)
– Definition: Substantive and procedural fairness; treat people equally and distribute AI’s benefits/burdens justly; avoid discrimination.
– Example: Biased hiring systems rejecting women due to skewed training data.
– Explicability (Group 4: Anoush, Imat, Barfia, Ermahan, Niloufar)
– Aim: Make reasoning and decision paths understandable (white-box where feasible); otherwise ensure traceability, auditability, and transparent communication of capabilities/limits.
– Examples: Recommender systems and LLM behavior; illustration using “Kyrgyzstanâ€� naming learned from user chats.
– Activity 2: Applying one principle—Prevention of harm in self-driving “trolleyâ€� scenarios
– Case setup: Tesla Autopilot encounters an unavoidable choice—hit a pedestrian or swerve into a tree (risk to driver). Question: How should “do no harmâ€� be operationalized?
– Group outputs and discussion:
– Least-harm calculus: Choose the option with lower expected harm (e.g., swerve into tree if airbags make driver injury less likely than pedestrian fatality).
– Protect the buyer/driver: Market logic—buyers expect their safety to be prioritized; system should prefer driver preservation.
– Design for anticipatory safety: Improve sensing/recognition (e.g., detect children vs adults), proactively reduce speed in higher-risk contexts, and extensively lab-test scenarios before market release; do not ship if not ready.
– Instructor synthesis: Three programmable policies exist—(1) dynamic least-harm rule; (2) always prioritize driver; (3) always prioritize third-party pedestrians (on grounds that drivers assume risk when choosing self-driving tech). Each aligns differently with “prevention of harmâ€� and raises unresolved questions of responsibility and ethical valuation.
– Meta-note: Classic trolley problem—no consensus answer; highlights difficulty of applying even a single principle consistently.
– Activity 3 (preview): Reconciling multiple principles in public safety surveillance
– Case: After a Brussels subway bombing, city considers facial recognition to match faces against local/international watchlists.
– Principle tensions:
– Autonomy: Mass surveillance undermines individuals’ ability to act without pervasive tracking.
– Prevention of harm: Public safety gains vs harms from false positives, chilling effects, and misuse.
– Fairness: Even if improved, bias risks persist; provenance/quality of watchlists matter.
– Explicability: Opaque matching pipelines and unclear data sources complicate accountability.
– Next steps: Read OECD AI guidelines; compare with EU principles (similarities/differences); prepare to amend or strengthen proposed rules to resolve conflicts.
Actionable Items
Urgent (before next class)
– Post readings:
– Upload OECD AI Principles (link/PDF) to eCourse; email to Elijah as promised.
– Repost a clean link/citation for the EU Trustworthy AI guidelines and a one-page summary of the four principles (to avoid copy/paste issues seen in class).
– Set expectations:
– Specify the deliverable for the EU vs OECD comparison (e.g., 250–300-word reflection or 1–2 slide comparison).
– Confirm whether students should come prepared to amend/tweak one existing proposal (e.g., CCPA amendment, FTC labeling, EU election rule) in light of conflicting principles.
Follow-ups/clarifications
– Verify and address the “AI minister in Albaniaâ€� anecdote with sources next session to reinforce evidence-based discussion.
– Ask Samira to clarify whether her proposal is for an EU AI transparency “task forceâ€� (and to outline mandate/scope); rectify any “tax forceâ€� transcription error.
– Invite Mohamed Omar to share draft amendment language for a “meaningful explanationâ€� right under CCPA as a model for the “amend existing rulesâ€� exercise.
– Record breakout group membership in LMS notes for participation tracking:
– Group 1 (Autonomy): Bekaim, Elijah, Gavin, Mohamed Omar
– Group 2 (Prevention of harm): Izirek, Amira, Safi, Wuti
– Group 3 (Fairness): Freshta, Samira, Banu
– Group 4 (Explicability): Anoush, Imat, Barfia, Ermahan, Niloufar
Pedagogical/assessment
– Short writing prompt (optional but recommended): Each student selects one self-driving harm policy (least harm, prioritize driver, prioritize third party) and justifies it using the EU principles (max 250 words).
– Plan a structured debate/case lab on post-attack facial recognition: students must propose safeguards or amendments that reconcile autonomy, harm prevention, fairness, and explicability, and justify trade-offs.
Homework Instructions:
ASSIGNMENT #1: Compare OECD AI Principles with EU Trustworthy AI Guidelines
You will read the OECD AI Principles posted for you and prepare a concise comparison with the EU’s four principles we worked through in class (respect for human autonomy, prevention of harm, fairness, explicability). This will help you connect our discussions (e.g., self-driving car dilemmas, facial recognition after a subway attack, and policy ideas like labeling synthetic media or amending CCPA) to concrete policy frameworks and get you ready for next class’s discussion.
Instructions:
1) Get the reading:
– Download the OECD AI Principles document posted on eCourse (the professor said it would be posted tonight; Elijah will also receive it directly).
– Re-skim the EU “trustworthy AIâ€� guidelines you were assigned over the weekend, focusing on the four principles we discussed: respect for human autonomy, prevention of harm, fairness, explicability.
2) Identify the core principles:
– List the OECD principles in your notes (e.g., human-centered values/fairness, transparency/explicability, robustness/safety/security, accountability, inclusive growth/well-being).
– List the EU four principles and the short definitions we articulated in class:
• Respect for human autonomy: human decision-making comes first; humans can override AI.
• Prevention of harm: AI should not cause physical, psychological, social, economic, or environmental harm.
• Fairness: no bias or discrimination; fair distribution of benefits/burdens.
• Explicability: decisions should be transparent, traceable, and understandable.
3) Map OECD ↔ EU alignment:
– Create a side-by-side mapping (bullet list or simple table in your notes) showing where each OECD principle best aligns with one or more EU principles (e.g., OECD “robustness, security and safetyâ€� ↔ EU “prevention of harmâ€�).
– For each mapping, write 2–3 sentences on:
• Overlap in intent.
• Any differences in emphasis (e.g., OECD’s “accountability� vs EU’s focus on explicability and human oversight).
• Scope differences (e.g., OECD’s framing of inclusive growth/well-being vs EU’s rights-based framing).
4) Note distinctive features and gaps:
– Identify at least two areas where the OECD adds something the EU document does not foreground (e.g., explicit “accountabilityâ€� as a principle; emphasis on inclusive growth/innovation).
– Identify at least one area where the EU’s framing is stronger or clearer (e.g., explicit naming of human autonomy as a principle).
5) Apply both frameworks to two scenarios from class:
– Scenario A: Self-driving car “trolley problemâ€� (Tesla autopilot choosing between harming driver vs pedestrian).
• In 5–7 sentences, explain how each framework would guide design/decision-making (e.g., EU “prevention of harm� and “explicability� require documented, testable decision logic; OECD “robustness/safety� and “accountability� require rigorous pre-deployment testing and clear assignment of responsibility).
• State what each framework seems to prioritize and where trade-offs or ambiguities arise.
– Scenario B: Post-bombing facial recognition in the subway.
• In 5–7 sentences, evaluate the proposal through both frameworks, noting tensions between “prevention of harm� (public safety) and “respect for human autonomy,� “fairness,� and “explicability� (privacy, bias risks, transparency).
• Specify minimum safeguards each framework would likely demand (e.g., bias audits, error-rate thresholds, human-in-the-loop review, purpose limitation, redress mechanisms).
6) Propose one policy “tweak� grounded in both frameworks:
– Draft one concrete improvement (5–8 sentences) to an AI-related policy discussed by classmates, such as:
• Amending CCPA to guarantee meaningful explanations for significant automated government decisions (link to EU “explicability� and OECD “accountability�).
• Requiring labeling and a public registry for AI-generated political content (connect to fairness, transparency, and prevention of harm to democratic processes).
– Explain how your tweak addresses conflicts between principles (e.g., balancing autonomy/explicability with safety) and how you would enforce it (audits, penalties, oversight body).
7) Deliverable format and preparation:
– Prepare 1–2 pages (roughly 400–600 words) of organized notes that include:
• Your principle mapping.
• Scenario analyses (A and B).
• Your policy tweak with justification.
– Bring these notes to the next class and be ready to discuss; keep them clear enough to share or upload if requested.
– Cite principle names precisely and include page/section references when possible.
8) Quality checklist before class:
– Your comparisons reference both frameworks explicitly.
– You address concrete trade-offs we discussed (e.g., trolley problem logic, mass surveillance risks).
– Your policy tweak is specific, enforceable, and clearly tied to at least two principles from each framework.
Due: Bring your notes to our next class meeting (Tuesday) for discussion.