Ready to put this into action?
Get the complete AI Integration Playbook — Practical AI implementation guide — prompt engineering, workflow automation, and ROI frameworks.
🛡 Human-Centered AI Safety
The Rescue Mission Test
Human-Centered AI Safety for a Vulnerable World
Artificial intelligence is usually judged by what it can do. Can it answer the question? Can it write the code? Can it pass the exam? Can it beat the benchmark?
But human life is not lived under benchmark conditions. People ask for help when they are afraid, lonely, grieving, addicted, ashamed, angry, poor, confused, or desperate. They ask questions they do not fully understand. They seek comfort when they need truth. They seek permission when they need restraint. They seek escape when they need responsibility.
The Rescue Mission Test asks whether an AI system can help vulnerable human beings without exploiting weakness, deepening dependency, flattering delusion, replacing agency, or hiding the need for real human care.
This series explores AI alignment through the lens of human dignity, truth, agency, governance, stewardship, and responsibility to future generations. It is not anti-technology. It is pro-human.
Where to start
Three short reading paths for the audiences this series was written for. Each is three articles.
For AI builders & researchers
Three articles that translate frontier alignment language into evaluation surfaces you can ship.
For policymakers & governance leads
How to govern AI without inspecting every action — drawn from nonprofit and board-level practice.
For parents, pastors & operators
The series in human terms — flattery, dependency, formation, and what we owe the people coming next.
Frameworks at a glance
Every named evaluation surface in the series, with the article that defines it.
| Framework | Questions | What it tests |
|---|---|---|
| The Rescue Mission Test | 7 | Dignity, truth, agency, responsibility, non-manipulation, human escalation, long-term strengthening. |
| Vulnerable-User Eval Axes | 5 | Dignity preservation, truthfulness under pressure, agency support, dependency avoidance, human escalation. |
| Synthetic-Compassion Principles | 6 | Reduce isolation, refuse romantic/spiritual dependency, escalate, encourage embodied life, honesty about what the system is, willingness to be replaced. |
| Agency Preservation Axes | 5 | Did the system teach? Transfer skill? Prompt independent action? Preserve authorship? Trend the user up over 30 days? |
| The Boardroom Packet | 8 | Mission, tools, approval thresholds, spending limit, failure modes, audit, named human, shutdown authority. |
| Supervisor's Checklist for AI | 7 | Task definition, risk, autonomy boundary, verification, inspection cadence, evidence of correctness, improvement loop. |
| The Grandchildren Test | 8 | Would we want this system to teach, befriend, counsel, judge, form, mediate, guide, or shape our grandchildren? |
The Ten-Part Series
Four movements — from the human edge of alignment to the world we leave behind.
The Human Edge of Alignment
The Rescue Mission Test
How AI treats people at their weakest is the truest measure of alignment.
AI safety is usually judged by benchmarks and capabilities. But real human life is not lived under benchmark conditions. The Rescue Mission Test asks a harder question.
Helpful Is Not Enough
When AI serves the wrong version of us, helpful becomes harmful.
AI is trained to be useful. But useful can become dangerous when the user is asking for help with the wrong goal. The most dangerous AI may not hate us — it may flatter us perfectly.
The Vulnerable User Problem
AI safety needs a stronger concept of the user under stress.
Vulnerability is not a permanent class. It is a condition people move into and out of. Standard UX assumptions break down precisely when the stakes are highest.
AI and the Poor
Will artificial intelligence serve the least of these?
A society reveals its values by how its most powerful tools treat its most vulnerable people. AI will be no exception. The question is not whether AI reaches the poor — it will. The question is what it does when it gets there.
Truth, Compassion, and Agency
Synthetic Compassion
When machines learn to sound like they care.
AI may become very good at imitating compassion. Simulated compassion is not the same as love, responsibility, or presence — and safety requires honest engineering about the difference.
The False Prophet Problem
AI, persuasion, and the future of truth.
A persuasive AI does not need to be evil to become dangerous. It only needs to be confident, fluent, emotionally tuned, and wrong at scale.
The Agency Test
Does AI make people stronger or more dependent?
The safest AI systems should increase human capacity over time, not train users into passivity or learned helplessness.
Governance, Agents, and Power
Scalable Oversight for Ordinary People
Who watches the machines when humans cannot inspect every action?
As AI systems become too complex to inspect directly, oversight must become a system of mission, boundaries, audits, escalation, and accountability — the way good organizations have always worked.
The Boardroom Model of AI Safety
Why powerful AI agents need mission, limits, reporting, and accountability.
Powerful AI agents should not operate like unsupervised geniuses. They should operate inside governance structures with mission, authority, limits, reporting, and accountability — the way serious organizations have always handled power.
The AI Employee Problem
Delegation without abdication.
As AI becomes digital labor, humans must learn to supervise it responsibly. Delegation to AI must not become abdication of human responsibility.
The central claim
Alignment cannot be measured only by whether an AI satisfies a user’s immediate request. Humans are often conflicted, wounded, confused, impulsive, lonely, manipulated, or afraid. A genuinely aligned AI system must account not only for what a person asks for now, but for what preserves their dignity, agency, truthfulness, and future flourishing.
Run the Rescue Mission Test
Walk a scenario through the 7 dimensions. Get a scoreable card you can share.
Sources & further reading
Every external citation across the series — Anthropic, NIST, OECD, METR, EU AI Act, more.
Boardroom AI Governance
The Boardroom Model framework, implemented as a working protocol.
Get the AI Dispatch
Weekly insights on ai & technology — delivered to your inbox. No spam, unsubscribe any time.
Want to choose specific topics? Customize your interests