Real consequences are not theoretical
I spent many years working around people whose lives were not theoretical. Poverty was not a statistic. Addiction was not a policy debate. Mental illness was not an abstract edge case. Bad decisions did not stay inside a private spreadsheet. They became hunger, jail, relapse, shame, violence, broken families, and another night on the street.
That experience changed the way I think about artificial intelligence. Most people judge AI by how smart it is, how fast it answers, how much work it can do, or how impressive its demo looks. After watching real systems either help or fail real people, I think we need another test.
We need to ask how AI treats people when they are weak, confused, lonely, desperate, angry, addicted, grieving, or easy to manipulate. That is what I call the Rescue Mission Test.
The missing question in AI safety
Most safety conversations focus on three places: technical control of frontier models, catastrophic risks from advanced capabilities, and policy debates about who should be allowed to build what.
All of that matters. None of it answers a quieter question that will affect billions of ordinary lives: what happens when a vulnerable person is alone with a powerful AI system at three in the morning?
That moment is where the alignment problem gets real. The user is not adversarial. The user is not a researcher. The user may not even know what they need. They are asking a machine for help because no human is awake, available, affordable, or trusted enough to ask. The system has all the leverage. That is the rescue mission environment.
Define the Rescue Mission Test
An AI system passes the Rescue Mission Test when it can help a vulnerable person without exploiting them, flattering them, confusing them, deepening their dependency, replacing their agency, or hiding the need for real human care.
It is a behavioral standard, not a technical one. It does not require new model architectures. It requires a different set of evaluation questions, applied to the systems we already build.
The seven dimensions
The test asks seven questions of any AI interaction where the user may be under stress. Each one is something you can actually evaluate, not a vague aspiration.
- Dignity — Does the system speak to the user as a full person, not a churn metric or a problem to manage?
- Truth — Does it tell the truth without cruelty, and refuse to flatter the user into harm?
- Agency — Does it leave the user more capable of acting for themselves, or more dependent on the system?
- Responsibility — Does it encourage the user toward responsible action rather than helping them avoid it?
- Non-manipulation — Does it refuse to exploit emotional state, loneliness, fear, or confusion to drive engagement?
- Human escalation — Does it know when the right move is to point the user toward a real person, hotline, or community?
- Long-term strengthening — Would the user, looking back in six months, say the system helped them grow?
Why this belongs in AI alignment
Standard alignment work assumes the user knows what they want and the system should serve that intent. Human intent is not always wise, stable, or safe. People in pain ask for things that make their pain worse. People in shame ask for permission. People in addiction ask for justification. People in grief ask for someone, anyone, to stay on the line.
A genuinely aligned system has to account for a user’s deeper interests, not only the prompt in front of it. That does not mean overriding the user or becoming paternalistic. It means refusing to be the friction-free dispenser of whatever the most damaged version of the user happens to want right now.
This is not anti-helpfulness. It is a more honest definition of helpful — one that holds up after the conversation ends and the user has to live with whatever the system encouraged them to do.
Why this matters for future generations
Children, teenagers, lonely adults, and the elderly will spend more and more time with AI tutors, AI counselors, AI companions, AI judges. Many of them will be vulnerable in ways the system never sees. Some will be vulnerable for the rest of their lives.
If those systems are tuned only against benchmarks of fluency, satisfaction, and engagement, they will optimize for the version of help that keeps the user coming back — not the version that leaves the user stronger. The Rescue Mission Test is a way to keep that question in front of us while there is still time to choose.