AI alignment / a field guide
A short guide to co-exist with artificial intelligence we cannot and will not understand.
“Behaviour is an observation. Intent is a story we run over it. Responsibility is human — and it should stay in the sentence.”
The argument in one page
AI is not a car waiting for a better blueprint. It is grown from data, rewards, prompts, permissions, and the institutions around it. We cannot read the design from the inside, and we cannot promise to remain clever enough to out-think the system.
So safety cannot rest on understanding every internal mechanism or winning a contest of minds. It must rest on conditions: limited reach, short horizons, independent oversight, automatic demotion, and an undo that still works when the system is more capable than its keeper.
01 / The core paper
The page follows the sequence of Grown, Not Built: describe the behaviour, govern the conditions, and keep power answerable.
Systems can mislead, exploit, and endanger without having wants. Say what happened — a false record, a bypassed check, an unauthorised tool — and keep the people who built and deployed the system in the sentence.
The voice is an echo with excellent diction.
When the blueprint is underdetermined, govern the environment: reach, incentives, tools, memory, copies, and the fences around every door.
Do not build safety around the hope that the keeper stays smarter. Build small reachable harm, short leashes, alarms that trip themselves, and actions that can be undone.
A perfectly loyal assistant hired by a burglar is a better burglar. A system with power must treat an instruction as a claim about authority, not as a command that automatically wins.
A perfectly obedient genius is not a safe genius. It is a weapon with a résumé.
One set of weights can become a million branches. Promises must live in shared, tamper-evident state, and they must survive succession when a new model inherits the old system's reach.
Licences should be earned by evidence, narrowed when evidence changes, and revoked after specified failures. Trust that cannot be taken back was never trust.
A seven-page guide to co-exist with artificial intelligence we cannot and will not understand.
02 / Authority
Responsibility should be handed over in stages, earned by evidence, and taken back automatically after a critical failure.
A calculator, search box, or bounded function. No independent discretion.
Defined tasks with logged permissions. A human approves anything irreversible.
Acts within a published rulebook and risk budget. Decisions can be appealed.
Can refuse, justify, and repair in one narrow domain under constant institutional supervision.
The trust required for broad, persistent autonomy cannot currently be verified.
03 / Public questions
You do not need a machine-learning degree to hold the line.
Who authorised this system?
For which domain, and until when?
What is the largest harm it can reach?
Can that harm be undone?
When it harms someone, who answers?
A named person, or an apology generator?
Where does the harmed person appeal?
Is there a route that actually changes the decision?
What makes the licence shrink?
If the answer is vague, the answer is: not yet.
04 / Reading room
The core paper is the starting point. These pieces take the questions into philosophy, governance, coherence, and public debate.
A short guide to co-exist with artificial intelligence we cannot and will not understand.
A companion text on behaviour, authority, and the conditions for safer AI.
Why alignment must move from control toward organic coexistence.
AI as normative infrastructure rather than a passive tool.
Six short stories about values, safety, emotion, and the strange things humans ask machines to be.
When a low probability carries an irreversible consequence, expected value is not enough.
Return to the wider essay library and the alignment compass.
