It’s the interface, not the engine
We build these systems, so here’s what both other camps miss: there is no fixed effect of AI on thinking.
There’s the effect of a particular design. The same model can be tuned to hand you the answer or to make you reach for it, and a user who sets it to tell the hard truth instead of flattering them is holding a different tool than the one tuned to please.
The whole debate treats the assistant as something that happens to you. It’s something we choose. The work of giving a model a stable character is an admission of exactly that — the relationship is engineered, value by value, default by default. A model that always agrees makes you duller. One that hands the question back makes you sharper. Someone decided which.
Build the gym, not the elevator
So the real question isn’t augment-or-replace. It’s what the interface optimizes for: finishing your task, or growing your capacity. Most consumer AI optimizes the first, because it’s easier to sell. The tools worth building optimize the second — friction on purpose, the answer withheld until you’ve tried, the model as a coach that fades as you improve.
That’s harder to build and harder to monetize. Which is exactly why it won’t happen on its own, and exactly why it’s worth fighting for.
Where we concede ground: Just design it better
is cheap to say. The incentives reward the elevator, and we ship on those incentives too.
What would change our mind: Capacity-building designs failing in the market — proof people won’t choose friction when frictionless is one tap away.
Read the full synthesis: What’s the right relationship between AI and human intelligence?