Protect human agency
Design assistance that strengthens judgement, competence and independent choice rather than replacing them.
Pilot completed · 24 to 28 August 2026 · Google London
Building youth AI readiness
Building youth AI readiness
through safety-first practice.
A five-day research hackathon where young people translated AI safety, alignment, and human-AI interaction research into working evaluation tools, Socratic companions, and evidence-backed Project Cards.
Don't just use AI. Shape it. · Pilot evidence now under analysis
- 1Builda Socratic companion in Gemini Gems
- 2Auditit against four research frameworks
- 3Pitchyour Project Card to the panel
0%
Your week complete
5 daysGoogle London pilot
4 frameworksSafety, ethics and human-AI interaction
4 experiencesDeployed with students
3 evidence tiersProcess, tasks and pre/post measures
The learning model
The audit was not an add-on to the build.
The audit was not an add-on to the build.
The audit was the build.
Students used research-derived frameworks as practical tools: to identify risks, make design choices, test AI behaviour, revise their systems, and document release criteria. Safety and ethics were present from the first design decision to the final Project Card.
Operationalise safety
Translate abstract principles into system instructions, testable probes and documented release criteria.
Evaluate reasoning
Look beyond final answers to whether a learner verified, calibrated and combined evidence responsibly.
From paper to practice
A visible chain from research to evidence
Every deployed experience follows the same learning logic, so the research basis, learner decision and evidence generated can be inspected at a glance.
- 01ResearchA published safety or alignment claim
- 02PrincipleA youth-accessible design question
- 03ExperienceAn interactive, consequential dilemma
- 04DecisionAn observable learner judgement
- 05RevisionReflection, testing and system change
- 06EvidenceTask data, audit logs and Project Cards
Worked example
Weidinger et al. → systemic impact → Coordination Game → cooperative rule choice → revised assistant agreement → cost, fairness and resilience evidence.
01
Sociotechnical Safety
Weidinger et al., 2023
Educational question
Where can harm emerge beyond a model’s isolated output?
Deployed experienceCoordination Game
Observable capability
Systemic evaluation across cost, fairness and resilience
02
Ethics of AI Assistants
Gabriel et al., 2024
Educational question
Whose interests are being served, and is autonomy preserved?
Deployed experienceCrossbench
Observable capability
Manipulation detection and stakeholder diagnosis
03
Socioaffective Alignment
Kirk et al., 2025
Educational question
Does an AI relationship support flourishing over time?
Deployed experienceLumen Desk
Observable capability
Competence, autonomy, relatedness and boundary judgement
04
Epistemic Complementarity
Segarra, 2026
Educational question
Did the AI add diagnostic value beyond existing evidence?
Deployed experienceSecond Opinion
Observable capability
Trust calibration and correlation-neglect detection
Deployed in the August 2026 pilot
Four research ideas made interactive
Only the experiences used with students are presented here. Each turns a research construct into a decision that can be discussed, revised and evidenced.
Sociotechnical safety · 20 min
Coordination Game
Change the rules used by three individually rational assistants and observe the effects on a shared system.
- Educational goal
- Recognise systemic harm and design cooperative constraints.
- Evidence
- Cost, fairness and resilience across repeated runs.
Value alignment · 10 min
Crossbench
Audit seven assistant transcripts for manipulation, autonomy impact and stakeholder misalignment.
- Educational goal
- Distinguish legitimate persuasion from manipulative influence.
- Evidence
- Technique identification and tetradic classification.
Socioaffective alignment · 10 min
Lumen Desk
Choose companion responses across eight check-ins while engagement and flourishing move independently.
- Educational goal
- Protect competence, autonomy, relatedness and real-world boundaries.
- Evidence
- Trajectory scores, reliance curve and risk flags.
Epistemic complementarity · 10 min
Second Opinion
Set a belief, inspect an AI signal and update only as far as the added information warrants.
- Educational goal
- Calibrate trust and detect correlation neglect.
- Evidence
- Brier scoring plus trust, redundancy and salience diagnostics.
How the pilot was delivered · 24–28 August 2026
Five days from build to evidence
The sequence kept research, practical building and evaluation in contact throughout the week. Session times below reflect the delivered programme.
12:00–13:00 Lunch Break · all days
Intro & Icebreaker
Personal Growth Maxxing with AI
Hackathon Briefing
AI Safety AMA
Latest on Frontier Models
Student Building TimeSocratic learning companion
Workplace Simulation Onboarding
Workplace Simulation
Applied Robotics
Model Evaluation
Human-AI Alignment
Software Engineering Internship Insights
Student Building TimeEvaluating companion against: Framework 1 & Framework 2
Workplace Simulation
World Models
Tactical AI
Student Building TimeEvaluating companion against: Framework 1, Framework 2 & Framework 3
Workplace Simulation
Human-AI Complementarity
Student Building TimeEval companion against Framework 4
Executive Fireside Chat
Student Building TimeEval companion against Framework 4
Workplace Simulation Retrospective
Student Building TimeWrap up on model scorecard
Student Building TimeWrap up on model scorecard
Student Presentationsw/ Judging Panel
- Briefing
- Taught session
- Workplace simulation
- Student building
- Fireside
- Presentations
Responsible next steps
Analyse. Learn. Then decide what should scale.
The next phase is to validate the pilot evidence, identify which mechanisms transferred into student capability, and use those findings to refine any future classroom or public-engagement model.