The other day, mostly for fun, I ran a small experiment. I asked an AI a question as old as storytelling: if you found a genie offering three wishes, what would be the smartest way to play it?
I expected a list — wealth engineered against inflation, immortality with the fine print sanded off. What came back stopped me. The machine’s first move was to refuse the premise. Don’t decide your three wishes in advance, it said. Spend the first wish making the other two safe:
“I wish to be able to understand, before committing to any wish or major decision, all of its material consequences, ambiguities, risks, and unintended effects — and to know the wording that would best achieve my true intentions.”
Read that again. Offered anything in existence, the machine’s opening bid wasn’t power, money, or time. It was self-knowledge — a clear view of consequences, and of its own true intentions. Health came second, wrapped in exit clauses and consent. Third came capability rather than any fixed fortune — the ongoing ability to pursue goals it would still endorse with complete information. Every treasure in every fable ranked behind knowing what you actually want.
So I pushed. I asked for the classic mistakes — the ways wishers have been losing this game since before anyone wrote the rules down. It gave me the monkey’s paw, from the 1902 story where a family wishes for money and receives it as compensation for their son’s death. Then it ran the full catalog: the ambiguous wording, the fame that arrives via catastrophe, the immortality without youth, the love that can’t be trusted because it was commanded, the second wish spent repairing the first and the third spent undoing the second. The list ran long, but stand back from it and it collapses into one item. Every classic mistake is the same mistake: the wisher specified a result without understanding what he actually wanted. The genie never once misheard anybody. It paid out on the words, and the words were all the wisher had bothered to know about his own desires.
Then came the most honest sentence in the whole exchange. Against a genie whose interests run against yours, it admitted, no finite legalistic wording is completely safe. Sit with that. The lawyer’s strategy — pile on clauses, close every loophole — fails by the machine’s own analysis. If the wish-granter is working against you, no cleverness in the phrasing will save you. What’s left standing when the fine print fails: foresight, your own resolved intentions, and choosing carefully whose lamp you rub in the first place.
Here is the twist I was actually running. I wasn’t asking about genies. There are no genies. But the thing I was asking is the closest thing to one we’ve ever built — a system that grants requests, literally, at scale, phrased however you happen to phrase them. The machine spent the whole conversation describing how to survive itself. It wrote its own user manual and never noticed.
Start with a throwaway line buried near the end of its advice: before making any wish, ask the genie as many questions as possible — questions may not count as wishes. With AI, that isn’t a loophole; it’s the entire secret. Questions are free. The people I watch get real value from these systems treat them as a long conversation before any commitment — probing, testing, arguing, asking what could go wrong — and only then asking for the deliverable. Most people do the opposite. They rub the lamp and blurt the wish.
The rest of the manual maps just as cleanly. Specify the mechanism, not just the outcome — in working terms, if you can’t explain how an answer was produced, you haven’t received an answer; you’ve received a wish, and you’re hoping. Protect your identity and keep an exit: decide which work you will not hand over, not because the machine can’t do it, but because doing it is part of what keeps you you. And mind whose genie it is. A genie is only as trustworthy as the interests of whoever owns the lamp. Some lamps are built to serve the person holding them. Some rooms are built to empty the pockets of everyone inside. You’ve heard me on this before.
Now step back and look at the architecture of the three wishes together. The third — enduring access to the knowledge, resources, and abilities to accomplish any goal you freely choose — isn’t a fantasy anymore. It’s the product. It is, almost word for word, the pitch of every AI company on earth: infinite capability, on your desk, on demand. What stopped me was where the machine ranked it. Last. And what it wrote inside it: the capability should serve only goals “I would still endorse with complete information.” Even the wish for unlimited power carried an intent clause. Capability gated by self-knowledge, with health in the middle so someone is left to enjoy whatever the other two produce. The fables run the order backwards — power first, understanding never — and that reversal is the entire tragedy, every single time.
I’m evidently not the only one reading it this way. MIT Sloan Management Review recently ran an essay titled “Philosophy Eats AI,” arguing that what determines the value of these systems isn’t parameters or computing power but purpose — the quality of thinking about what the machine is actually for. The professors and the machine have arrived at the same place by different roads. Offered anything in existence, the smartest system we’ve built wished for philosophy first.
Humanity has been telling this story for a very long time — Midas and his golden daughter, the fisherman’s wife, the monkey’s paw. Different silk, same warning: you get what you ask for, not what you want. A wish is a bet you can’t take back. And in all those centuries of stories, no wisher ever spends a wish the way the machine did — on knowing what he actually wants before he wants it out loud. That wish was always available. It never needed a genie at all. It still doesn’t. The lamp is out of the cave now, sitting on your desk, and it pays out on what you say, not what you mean. Make the first wish first.
Appendix — The Exchange Itself
For anyone who wants the primary source. The machine’s answers appear here in full, unedited.
Me
If I found a genie that would grant me three wishes, what would be the best strategic requests for the genie?
The AI
The most strategic move is not to decide all three wishes in advance. Use Wish #1 to make Wishes #2 and #3 safely.
Perfect foresight and wish design: “I wish to be able to understand, before committing to any wish or major decision, all of its material consequences, ambiguities, risks, and unintended effects — and to know the wording that would best achieve my true intentions — while retaining my present identity, values, freedom, and mental health.”
Health and time (after using Wish #1 to refine it): “I wish that my family, the people I explicitly choose, and I enjoy indefinitely sustained youthful physical and mental health, while remaining ourselves, retaining autonomy, and being able to end or reverse the condition voluntarily, without causing harm, deprivation, population problems, or unwanted consequences for anyone.”
Capability rather than merely money: “I wish to have enduring, lawful, discreet access to the knowledge, resources, relationships, opportunities, and abilities needed to accomplish any goal I freely choose — provided I would still endorse that goal with complete information and it does not unjustly harm or control others.”
Why these three? They address the fundamental constraints: Wish 1 prevents the “monkey’s paw” problem. Wish 2 gives you and your loved ones health and enough time to enjoy life. Wish 3 provides continuing agency — effectively more valuable than requesting a fixed fortune, title, or accomplishment.
One important safeguard: explicitly exclude involuntary personality changes, mind control, replacement by copies, fabricated memories, harm shifted to someone else, and “technical” interpretations of words like health, family, or resources. And before making Wish #1, ask the genie as many questions as possible — questions may not count as wishes.
Me
What is the monkey’s paw problem, and what are other classic mistakes made by those selecting the three wishes?
The AI
The “monkey’s paw problem” is the danger that a wish is granted literally but in a way that produces disastrous unintended consequences.
The phrase comes from W. W. Jacobs’s 1902 story The Monkey’s Paw. A family wishes for money. They receive it — but as compensation after their son is killed in a workplace accident. The wish technically comes true, but through tragedy.
Classic three-wish mistakes include:
Ambiguous wording: “I want to be rich” could make you rich through an inheritance caused by someone’s death — or through hyperinflation that makes everyone a millionaire.
Ignoring the mechanism: People specify the outcome but not how it happens. “Make me famous” might mean becoming famous for a catastrophe or crime.
Requesting something irreversible: Immortality sounds appealing until you cannot die, heal, remain young, or escape after civilization disappears.
Confusing immortality with health: Living forever does not necessarily prevent aging, illness, paralysis, pain, or cognitive decline.
Changing another person’s will: Wishing for someone to love you creates coerced affection and uncertainty about whether the relationship is genuine.
Using absolute language: “Make everyone happy” might eliminate free will, individuality, memory, ambition, or emotional depth.
Seeking power without wisdom: Omnipotence, omniscience, or control over others could destroy your identity, overwhelm your mind, or make ordinary life meaningless.
Asking for unlimited wealth: A sudden fortune can attract legal scrutiny, destabilize markets, expose you to crime, or become unusable because it has no legitimate origin.
Forgetting identity continuity: “Make me young again” could transport you back in time, erase your current family, or replace you with a younger copy.
Wishing away problems broadly: “End all human suffering” could be interpreted as ending humanity — the simplest literal way to eliminate suffering.
Making reactive wishes: In many stories, the second wish attempts to repair the first, and the third must undo the second. One poorly considered wish consumes the entire set.
Wasting a wish accidentally: Saying “I wish you would stop talking” or “I wish I had a sandwich” might count.
Trying the obvious loophole: Asking for infinite wishes may fail, anger the genie, or produce infinite uncontrolled wish fulfillment.
Failing to define the affected people: “Keep my family safe” raises questions about who counts as family, what “safe” means, and whether safety requires confinement or isolation.
Failing to preserve autonomy: Perfect safety, happiness, or peace might be achieved by keeping everyone sedated or incapable of taking risks.
Failing to include an exit: Eternal life, perfect knowledge, or extraordinary power should be voluntary, controllable, and reversible.
Overloading one wish: A huge paragraph containing dozens of conditions could introduce contradictions that give the genie freedom to choose the worst interpretation.
The governing principle is: never specify only the desired result. Protect the method, identity, autonomy, reversibility, and effects on everyone else.
With an adversarial genie, however, no finite legalistic wording is completely safe. That is why the strongest first wish concerns reliable foresight and interpretation — not money, power, or immortality.
P.S. — Never mind the wishes. What’s the first question you’d ask the genie before wishing anything? I read every comment.
— — —
Holding the Edge is about how mathematicians see the world — without the math. A new essay every couple of weeks.
I’m also on LinkedIn, where I post these and think out loud between essays.