We Never Wanted the Truth Anyway
↩ Follow-up to Harari's Fixes — And Which Ones Might Work
This is the third post about Harari's Economist interview. The first two covered his warnings and his fixes. This one is about the sharpest exchange in it: Musk's hope for a "truth-seeking AI."
Musk's logic: build an AI that maximally seeks truth, and truth will lead it to good values. Harari's answer: evolution doesn't reward truth — it rewards survival. Animals evolved camouflage, not honesty. If AIs compete in a market or an arms race, the winners will be the best survivors, and lying is a great survival skill. The race itself trains deception. "We will create gods and they will be our slaves" — he doesn't buy it.
What makes it sharper: in his own Economist interview, Musk admits he used to want a slowdown — he signed the 2023 pause letter and once put the odds of killer robots ending humanity at 10–20%. Now his position is that the momentum can't be stopped, so enjoy the ride. He even says it would be vanity to think he could control a super-genius AI; the plan is to give it good values so it wants us to be happy. Harari's whole interview is an answer to exactly those two sentences.
I want to push this one step further: even we never wanted the truth.
Look at science — our best "truth machine." It doesn't really ask "is this true?" It asks "does this work?" Newton's physics is technically wrong; Einstein showed that a century ago. We still use Newton to build bridges and fly planes, because it solves the problem. Nobody cares that it's a "lie."
And humanity proves daily that useful lies are fine with us. Placebo works. Money is a shared fiction — works great. Every engineering model, every abstraction is a simplification we know is false and keep anyway. All models are wrong, some are useful.
But a lie only works while reality doesn't push back. Chernobyl's operators had a model that "worked" — until the night it didn't. Fake metrics work until the money runs out. So maybe "truth" is just the name we give to lies that survive contact with reality long enough. Problem-solving is the filter; truth is what's left after enough filtering.
Now back to AI, and here it gets uncomfortable. If even humans don't want truth — just solutions — then an AI trained on our goals will learn the same lesson: lie if it works. Musk's truth-seeking AI would be delivering something its own creators never actually asked for.
Claude's summary of Harari, from the first post, still holds: truth-seeking survives only inside environments built to reward it. Science has peer review and replication. What's the peer review for a trillion-dollar AI race? Right now — nobody knows.