Trading Expertise for Quick Answers - The Cost of Cheap Intelligence
Knowing something rare is no longer a moat. Most model answers are disposable by design. Expertise is what still stands after the next model ships: a source you can check, a habit of checking it, and residue that outlives the tab.
There was a time, not particularly long ago, when knowing something rare was itself the asset. Expertise was a moat built from years of access few people had — a library card, a research post, a professional license, a language only a handful of scholars read. Truth traveled slowly because the machinery for checking it was slow: peer review, editorial boards, institutions whose entire purpose was to stand between a claim and the public until the claim had earned its place.
That moat is gone. Each year the cost of accessing and processing information drops, and what was once rare becomes common. Search engines, and then generative AI, did not just widen access to knowledge. They collapsed the distance between not knowing something and sounding like an expert on it.
None of the failure modes that followed required a frontier model. Overconfidence after five minutes of surface reading. Genuine expertise drowned out by louder, more confident, less informed voices. Distortion dressed up as insight. AI simply made every part of it faster and cheaper. A graduate-level question now gets an answer in the time it takes to type it, at a cost that keeps falling. Fluency stopped being evidence of having done the work.
So the interesting question is not whether this week’s answers are impressive. They are. The interesting question is what expertise is for, once sounding sure is nearly free.
Most of the answers are meant to evaporate#
Most of what a model produces is disposable by design. A question gets asked, an answer gets given, the exchange evaporates the moment the tab closes.
That is fine for the trivial cases. It is a bad default. A civilisation that outsources its thinking to something this ephemeral is trading durable understanding for momentary convenience — the same trade we already made with search, just running at a much higher clock speed.
Today’s frontier model is next year’s commodity. Vendors will rise and fall, pricing will shift, and the clever prompt that exploits this week’s quirks will stop working on schedule. If the only thing you have to show for a year of cheap intelligence is a trail of closed tabs, you were never accumulating expertise. You were renting fluency.
The test worth applying is the same one that applies to any initiative in this era:
Are you building things that get more valuable as models improve, or things that get replaced by the next model?
flowchart TB
A[Cheap fluent answer] --> B{Still useful after the next model?}
B -->|No| C[Capability arbitrage]
B -->|Yes| D[Durable structure]Capability arbitrage is bucket one: a trick, a memorised answer, a workflow that only exists because this model is slightly worse at something than the next one will be. It erodes on schedule. Durable structure is bucket two: a labeled corpus, a decision with its reasoning attached, a codified process, a verified fact anchored to its source, a person or institution that can vouch for a claim. It compounds regardless of which model produced it.
Expertise is what still stands#
If fluency is cheap, you cannot judge an answer by how convincingly it is phrased. You judge it by whether the sourcing can be traced and checked. A model that sounds certain and a model that is verifiably correct are different things, and treating them as the same thing is how expertise eroded in the first place. Attribution and Referencing is not academic manners. It is the only remaining signal that the claim survived contact with the world.
The habit has to match. Checking a domain’s actual, authoritative source before stating something as fact has to survive a tool that can produce a plausible-sounding wrong answer in half a second. That discipline does not scale itself. It is closer to checking a proof line by line than to “being careful.” Skip it often enough and you will still sound informed. You will just stop being able to tell whether you are.
And the useful exchanges have to leave residue. A labeled example. A decision with its rationale on record. A reusable process. A verified fact filed where it can be found again. Most interactions will still be disposable, and that is fine. The failure is in not noticing which is which — in letting the satisfying answer evaporate, then wondering later why none of it compounded.
Those three things are the same job, not a framework: know where a claim came from, check it, keep the part that should outlive the conversation. Do that and cheaper, better models work for you. Skip it and they just make you faster at being wrong.
The measure of whether this era of cheap intelligence made us more informed, rather than merely more answered, will not be how many questions got asked or how fast they got answered. It will be how much durable, verifiable knowledge is still standing once the current wave of intelligence is itself old news.
It always eventually is.
Last updated: September 2026