AI changed both the supply and the demand of knowledge. The supply has multiplied — any motivated reader can produce a plausible-looking summary of any field in minutes. The demand has shifted with it. The skill that used to matter was finding good information. The skill that matters now is evaluating it, synthesising across sources, and noticing when something only sounds correct.
This is not a lament. The new shape of knowledge work is genuinely better for people who learn to use it well. But it requires habits most readers never had to develop, because the old constraints did the work for them. When books were expensive and journals were slow, the filter was upstream. The reader inherited the curation. That filter is gone now, replaced by a stream you have to filter yourself.
The three shifts
Volume is up. Signal is constant. The total amount of accessible writing on almost any topic has at least doubled in the last three years and the ratio of useful to forgettable is the same as it was. A reader who used to manage by reading three things a week now has thirty things to choose between, and most of them weren't worth writing.
Synthesis matters more than memorisation. The marginal value of holding a fact in your head has fallen sharply — you can retrieve it on demand. The marginal value of recognising the same idea expressed differently across two sources, noticing they contradict each other in a meaningful way, and resolving the contradiction has gone up. This is what reading deeply trains. Reading widely does not.
Evaluation skill compounds. Anyone can produce a clean-looking summary, including AI agents with no real grounding in the underlying material. Telling apart a summary written by someone who understood the source from one written by someone who didn't is the highest-leverage reading skill in this environment. It is also the hardest one to teach, because the only way to develop it is to spend a long time being wrong and then noticing.
Three loops worth practising
Source verification. When a piece of writing makes a specific empirical claim — "studies show," "research has demonstrated," "X has been linked to Y" — go find the underlying study. Read its abstract, at minimum. The disconnect between what summaries say and what the cited research actually claims is enormous, and it is consistent across human-written and AI-written summaries. The summary almost always overstates the strength of the evidence. Doing this twenty times across your field of interest will permanently change how you read everything else.
Cross-source synthesis. Pick a question you care about. Read three serious pieces by people who disagree. Notice exactly where they disagree — not the headline-level disagreement, but the implicit assumption each is making that produces their conclusion. Most public disagreements turn out to be disagreements about which question to ask, not which answer is correct. This is the cheapest way to learn what is actually being debated in a field versus what is merely surface noise.
Productive disagreement. When you find a position you agree with, ask an AI to argue the strongest possible case against it. Read that case carefully. If your view survives, you understand it better. If it doesn't, you have learned something. This trains the habit of holding a position lightly enough to update on real evidence and tightly enough to defend it against weak attacks. AI models are very good at this exercise specifically because they are not personally invested in the answer.
What changed about evaluation
The traditional signals of credibility — institutional affiliation, peer review, length of citation list — are less informative than they used to be. Papers come out of universities with hallucinated citations now. Long reference lists can be padded by anyone with an LLM and ten minutes. Institutional affiliation tells you something about a researcher's incentives, not the quality of their reasoning.
The signals that still work are harder to fake. Does the argument acknowledge the strongest counter-argument? Are the claims appropriately hedged? Does the writer distinguish what they have shown from what they are speculating about? Do they cite where their data comes from, and does the data actually exist? These are the things a reader has to check directly, because no upstream filter is going to check them for you.
If a piece doesn't acknowledge that it could be wrong, it is probably wrong in ways the author hasn't noticed. This is true of human writing and AI writing in equal measure.
The role of judgment
None of this argues against using AI to learn. The opposite — used carefully, an AI model is the most patient tutor most readers will ever have access to. It will explain a concept ten different ways without losing interest. It will follow a chain of "but why" questions to depths no textbook would bother with. It will challenge a half-formed thought into a sharper one.
What it cannot do, and will never be able to do for you, is decide which questions are worth asking. That is the part you have to bring. The reader who can articulate a precise question gets a far better answer than the reader who can articulate only a vague topic. The reader who can recognise a half-true answer can use AI productively. The reader who can't will be misled at scale.
The new shape of knowledge work rewards a small number of disciplined habits — verification, synthesis, deliberate disagreement, calibrated trust — and punishes their absence more harshly than the old shape did. The good news is that these habits are learnable. The work, if you want to do it, is largely the same work serious readers have always done. It just has to be done deliberately now, because no one else is going to do it for you.