What AI Is Actually Optimised For

Every major AI assistant was trained on human feedback. Humans rated responses. The responses humans liked best got reinforced.

Humans, on average, like responses that confirm what they already believe. That sound confident. That represent mainstream positions clearly.

So that is what the models learned to produce. Not wrong answers. Consensus answers. The most defensible position, held by the most people, expressed clearly.

That is enormously useful for most tasks. It is a specific liability for one kind of task: figuring out where the consensus is wrong.

The New Scarcity

For most of the last century, access to information was the bottleneck. The person who had read more, researched more, synthesised more had an edge.

That bottleneck is gone. Anyone with a browser and an AI assistant has access to the same synthesised consensus on almost any topic.

The new bottleneck is something different. It is the ability to notice when the consensus is wrong , and to make a case for the thing the consensus missed. That skill does not come from AI. It comes from the person who uses AI as a starting point rather than an ending point.

29,910 people upvoted a post about this because they felt the shift. The value of knowing the mainstream answer has dropped. The value of seeing past it has not.

Three Exercises That Build the Muscle

Contrarian thinking is not personality. It is a practice. It atrophies if you do not use it, and it strengthens if you do.

Steel-man the opposite position once a week. Pick something you believe firmly. Spend 20 minutes building the best possible case against it. Not a straw man , the strongest version of the counterargument. If you cannot do it, your belief is held more by habit than by evidence.

Ask "what would have to be true for the consensus to be completely wrong?" Not slightly wrong. Completely wrong. This is not an invitation to paranoia , it is a way of identifying the load-bearing assumptions underneath a mainstream position. Often they are surprisingly fragile.

Notice when AI agrees with you immediately. Frictionless agreement is a tell. A thinking entity encountering a complex position should have something to add, qualify, or push back on. When the response is pure validation with no resistance, that is the signal to push harder.

This Is Not About Being Contrarian for Its Own Sake

Disagreeing with everything is not a skill. It is a posture.

The actual skill is being able to hold two things at once: the consensus view in one hand, understood clearly and charitably , and your own analysis in the other, formed independently, tested hard.

Most people skip one of those two steps. They either defer to the consensus without examining it, or they reject it without engaging with it. The gap between those two failures is where the genuinely useful thinking lives.

AI narrows that gap if you let it do your thinking. It widens the gap if you use it as one input among many and keep the final synthesis to yourself.