A summary is lossy compression, and the loss is not random. It keeps the headline and discards the caveats, the effect size, the boundary conditions, the sample, the "we assumed X" — which is to say it discards exactly the information you need to decide whether the claim is true, applies to your case, or was overstated at the source. As AI drives the cost of summaries to zero and everyone converges on the same compressed version, the person who goes to the primary source is reading a different, better book. My bet is that the edge between them is widening, and that it compounds.
This is not a complaint about attention spans. It is an argument about where information value is moving — and one you can act on this week.
The compression is optimized for the wrong thing
Every lossy codec throws away information under an objective. JPEG discards the high-frequency detail your eye is bad at noticing, so the file shrinks and the picture still looks right. The discard is deliberate and tuned to a fidelity metric: preserve what the viewer will perceive, drop what they won't.
A summary is a codec too, and its fidelity metric is preserve the gist a reader will remember and repeat. That objective is indifferent to the conditions under which the gist holds. And judgment is nothing but conditions.
Write any real claim in its full form and the problem is immediate. A scientific result is not "E is true." It is "effect E holds for population P, under conditions C, with magnitude M, and uncertainty U." The summary keeps E. P, C, M, and U are the first things to go, because they are less quotable, cost more words, and blunt the punch. But P, C, M, and U are the entire content of judgment. They are how you know whether E is load-bearing for your decision or an artifact of someone else's setup. The codec strips precisely the bits with the highest decision-value, because those bits are also the least memorable and the least fun to repeat.
So the loss is not random noise around a faithful core. It is a systematic deletion of the machinery of doubt.
The abstract and the results table tell different stories
I do this for a living, so let me show the mechanism rather than assert it. In my own field — epigenetic aging clocks, lncRNA, microbiome causal inference — the gap between a paper's abstract and its actual results table is not an occasional embarrassment. It is the normal state of a paper.
Start with the fact that the abstract is already a lossy summary, written by authors whose incentive is citations and whose reward function favors a clean headline. An AI summary of the paper is then a compression of a compression — a codec applied to a codec's output, each pass optimized to keep the punchline. Meanwhile the ground truth sits untouched in the methods and the results table, where nobody is trying to make you feel anything.
Take a claim of the form "intervention reverses epigenetic age." The abstract says it cleanly. The methods say: n was small, the work was in vitro, it was one cell type, the clock was never validated on that tissue, and the measured reversal sits inside the clock's own error band. Every one of those is fatal to the headline, and not one of them survives to the summary. The caveat that kills the claim lives in the methods — the section a summary is structurally guaranteed to compress hardest.
The effect-size version is even cleaner. "Significantly associated with mortality" is true and useless when the hazard ratio is 1.05. Statistical significance is about whether an effect is distinguishable from zero given the sample; it says nothing about whether the effect is large enough to matter. The summary keeps "significant" and drops "1.05," because 1.05 is not a headline. You are left believing something real was found, and in the sense that would change a decision, nothing was.
The causal-inference version is the one I find most instructive, because it connects to a deeper point: a model of correlations is not a model of the world, and fluency is not truth. An observational microbiome study reports an association, tucks its adjustment set into a supplement, and concedes residual confounding in a single sentence of the discussion. That sentence — "we cannot exclude reverse causation" — is the one that should stop you from acting. It is also the one no summary on earth will retain, because it is the least quotable line in the paper and it argues against the paper's own headline. The overclaim gets laundered into fact one compression pass at a time, and the audit trail back to the doubt is exactly what got deleted.
When summaries are free, the source is where the alpha is
Here is why this is getting worse, and why that is good news for you.
An information edge is an asymmetry. It exists only in the gap between what you know and what the people you compete or trade with know. When digesting a source into a usable summary was itself expensive — when you had to read the thing and think — few people held even the compressed version, and holding it was an edge. That era is ending. The summary is now free, instant, and identical for everyone who asks the same model the same question.
Follow that through. If the compressed representation is free and universal, the median understanding of any topic converges to the summary's level. And anything priced into the median is, by definition, not an edge. The whole remaining spread — the entire uncontested part of the distribution — is the delta between the source and the compression. That delta used to be small relative to the cost of acquiring the summary at all. Now the summary costs nothing and the delta is the only thing left to own.
So the marginal value of reading the source rises as summaries proliferate. Picture a market where ninety-five percent of participants operate on the same free compressed representation. Your edge is not your cleverness on top of the summary; everyone has the same summary and the same models to reason over it. Your edge is the information the summary threw away — the boundary condition, the effect size, the assumption — because that is the one input no competitor sharing your feed has seen. Summaries commoditize the middle of the knowledge distribution and, in doing so, hand a monopoly on the tails to whoever still reads primary.
Generation is cheap; checking it against the source is the work
There is a reason this asymmetry doesn't close on its own. AI is spectacular at producing the summary, the plausible take, the confident well-formed paragraph. It is structurally unable to do the one act that creates the value: hold the generated claim up against the thing it claims about and check.
The automated scientist is a category error for essentially this reason — generation and verification are different operations, and only one of them makes contact with reality. Summarizing is generation. It is cheap, and getting cheaper by the quarter. Verifying a claim against its source is the expensive half, because it requires touching ground truth, and ground truth is the thing a language model does not have and you do. Reading the source is verification. That is why the task doesn't automate away and why the edge doesn't get arbitraged out from under you — the machines are flooding the world with generation and leaving the verification undone. The undone half is the whole opportunity.
You can't read everything — and that's not the skill anyway
The obvious objection is correct: life is short, the literature is infinite, and no one can read every source primary. Good. Treating primary reading as a monastic vow to distrust all summaries is not rigor, it is paralysis, and it would make you slower than the people running on the free version.
The actual skill is triage — knowing which few sources repay a deep read and compressing the rest without guilt. Read primary when the claim will drive a decision whose cost of being wrong is high; when the claim is surprising or too convenient, because striking-ness is what survives compression and is therefore the leading indicator of an overclaim; or when you are about to repeat it publicly and put your name behind it. Everything else, the summary is fine — use it and move on.
This maps straight onto operating, not just research. Building Kommerce, I learned not to trust a dashboard's headline metric — say, a clean cash-on-delivery acceptance rate. The aggregate is a summary, and it hides the exact failure mode I need: which cities, which couriers, which reason codes are dragging it down. So I go to the raw delivery-attempt records. The number on the dashboard is E; the decision lives in the P and C the dashboard compressed away. Same structure for a founder reading the actual churn-cohort data instead of the smoothed line, the real indemnity and liability clauses instead of counsel's one-line email, the competitor's actual filing instead of the press write-up. In every case the summary keeps the headline and deletes the thing you would have acted on.
The one reflex to build
Before you repeat a striking claim, trace it to its origin. The claim that made you want to act or forward it is, precisely because it is striking, the one most likely to be a laundered overclaim — striking-ness is what the codec is tuned to keep. One click to the source, thirty seconds on the results table or the raw data, before it leaves your mouth. Make "I read the actual paper, the actual data, the actual contract" a normal move rather than a heroic one, and it compounds: every source you read builds a model the summary-readers don't have, and models stack on each other in a way headlines never do.
Everyone has the summary. Almost no one has the thing. That gap is your edge, and it widens a little every time a model makes the summary cheaper.