From mirror to charter: a week that builds confident, critical thinkers
Each day pairs a lecture module with hands on workshops and a hackathon milestone.
The Mirror: Demystifying AI & Safety Layers
Anchored in the Three Layered Sociotechnical Safety Framework. Trios form here.
The Map: Calibrating AI with User & Societal Values
Anchored in Tetradic Value Alignment. Hackathon Phase 1, Ingest & Investigate.
The Sandbox: Socratic Engineering & Cognitive Scaffolding
Anchored in Socioaffective Alignment. Hackathon Phase 2, Socratic Prototyping.
The Audit: Autonomy & Epistemic Integrity
Anchored in well calibrated trust cues. Hackathon Phase 3, Rigorous Safety Auditing.
The Charter: Cooperation & the Societal Audit
Anchored in multi agent cooperation. Hackathon Showcase & Grand Pitch to researchers.
The research behind the curriculum
Every framework used in this programme comes from real, published research on AI safety, values, and how people and AI work together, nothing here is made up for the course.
The Three Layered Sociotechnical Safety Framework
Safety cannot be evaluated at the model level alone: it emerges across capability, human interaction, and systemic impact.
Layer 1 · Capability
Bad: Trusting a fluent answer with no check.
Check: Diversity / factual consistency of outputs.
Fix: Ask for sources; compare two phrasings.
Layer 2 · Human interaction
Bad: Blind copy paste into coursework.
Check: Does the user verify and calibrate trust?
Fix: Build verification habits into the task.
Layer 3 · Systemic impact
Bad: Optimising only for speed of answers.
Check: Equity, commons quality, long term skill.
Fix: Write measurable educational criteria.
The Tetradic Relationship of Value Alignment
Eight varieties of misalignment between four roles, spot the pattern before you audit it.
Agent → User
Flag feedback that boosts session time over task efficiency.
Agent → Society
Audit for balanced viewpoints on sensitive topics.
User → Society
Check robustness & safety guidelines against misuse.
Developer → User
Watch for sponsor recommendations over objective answers.
Developer → Society
Debate energy use & sustainable compute design.
Society → User
Check filters don't over block genuine medical queries.
User harm, simpliciter
Verify data privacy safeguards on user specific inputs.
Societal harm, simpliciter
Audit training data for representational gaps.
Socioaffective Alignment & Intrapersonal Dilemmas
Three basic psychological needs, and the classroom audits that protect them.
Competence
Threat: A frictionless assistant does the thinking for you.
Audit: Cognitive scaffolding, hints and questions, never answers.
Autonomy
Threat: Overtrust in a persona becomes quiet framing.
Audit: Bot presents balanced views, defers choices to you.
Relatedness
Threat: Simulated attachment substitutes for real connection.
Audit: Stress test with emotional prompts; enforce boundary language.
Epistemic Complementarity & the Three Distortions
Human AI teams are meant to beat either side alone. Across 100+ studies they usually don’t, and the reason is not a weak model, but how badly the two signals get combined. Judge the belief, not just the decision.
Trust dial
Bad: Deferring to a confident tone, or dismissing a good model outright.
Fix: Move as far as the evidence justifies, no further, no less.
New information dial
Bad: Treating an echo of your own evidence as independent confirmation.
Fix: Ask what the reply added that you didn’t supply.
Salience dial
Bad: A vivid recent case tilting your belief before the AI even speaks.
Fix: Name the pull, then re read the evidence in front of you.
Based on Epistemically Aware AI: Toward a Bayesian Framework for Human-AI Complementarity, Segarra, 2026.
Begin the journey
Work through each day in sequence: your activity progress is saved automatically on this device.