The web has had three acts. In the first one, through the late 1990s, a few people produced and everyone else consumed. Publishing took a server, some HTML, and patience with the sound of a modem negotiating with a phone line. The asymmetry was obvious and nobody pretended otherwise. In the second act, blogs and social media arrived and anyone could produce, which is about when Tim O’Reilly gave the thing a name (yes, I’m talking about Web 2.0). This was a real democratization, and I say that as someone who has spent a career studying how people look for information. Voices with no path to an audience finally got one. Sure, the comment sections were a warzone and your uncle discovered political commentary, but the trade was worth it. Production and consumption stayed in rough balance, mostly because writing still cost you an evening.

We are now in the third act, and generative AI has made production close to free. A report, a thread, a paper, a newsletter: one prompt and thirty seconds. So what changed? Everything on the supply side, and nothing at all on the demand side.

I still read at about 250 words a minute, the same rate I managed in graduate school, and I still have twenty-four hours in a day, some of which I insist on spending asleep. Whatever GenAI has done to the supply of text, it’s done nothing to the thing that actually limits knowledge, which is a human being paying attention long enough to understand something. We automated the easy half of the loop and left the hard half exactly where it was.

Don’t get me wrong. I’m not writing this as someone standing outside the building throwing rocks at it. I build these systems, I run a lab full of people making them better, and I use them every week. That’s precisely why the asymmetry bothers me.

I recently came across a story in The Washington Post about how a University of Chicago’s professor wrote 200 papers in a year. I can’t even read that many papers in a year! Now, this may seem like an outlier, but we are starting to see community-wide trends of such behavior.

Look at academic conferences. AAAI received fewer than a thousand submissions per year for its first three decades. For AAAI-26, the main technical track got nearly 29,000, with roughly 23,000 surviving to review, about double the previous year, and handling that took a program committee of more than 28,000 people, close to triple the year before, plus the largest AI-assisted review pilot the field has run, which produced a machine-generated review for essentially every paper. So we have arrived at a place where papers are increasingly written with AI help and then reviewed with AI help, while humans stand nearby holding a clipboard.

I have reviewed for these venues for twenty years, and I have been on the other side too, emailing colleagues to beg them to take three more papers. The conclusion is hard to avoid: a great many people want to publish, and almost nobody wants to read. The reviewing crisis is not a logistics failure. It’s a revealed preference.

Is all this volume at least giving us better work? The evidence isn’t cooperating. arXiv moderators rejected around 4% of submissions for years, and that number has since climbed to something like 10 to 12%, with the surge starting in early 2025 in computer science and spreading outward from there. arXiv stopped accepting computer science reviews and position papers unless they had already passed peer review, then added a one-year ban for authors who submit visibly unchecked model output. Fabricated citations, the easiest symptom to count, have risen roughly twelvefold since 2023. One preprint service simply stopped taking submissions because most of what arrived was too poor to host. A study in Organization Science put it about as bluntly as a journal can: the incentives are pushing us toward more research rather than better research.

In fairness, some of this growth is real. There are more AI researchers today than a decade ago, and more people doing good work produce more good work. But that defense doesn’t touch the reading side of the ledger. Whether the flood is “slop” or sincerity, nobody has time to drink it.

So what actually worries me here? Two things, and the second more than the first. The first, the obvious one, is shallowness. When there is always more to skim, skimming becomes the rational strategy, and we end up a society that’s heard of everything and understood almost nothing. Depth means staying with one difficult thing past the point of discomfort, which is exactly the behavior an endless feed punishes. The second is loss. Somewhere in those 23,000 submissions is the paper that would have changed how I think, and I’ll probably never see it, because good work doesn’t automatically float and at this volume the filters we lean on, meaning reputation, venue, and word of mouth, drown along with everything else. We’re not only producing garbage. We’re burying treasure underneath it.

So what do we do? I want to name the alternative, because unnamed ideas don’t travel.

Generative AI adds. What we need alongside it is “Subtractive AI”: systems judged by how much they take away. Michelangelo supposedly described sculpting as removing everything that is not the statue, and that is the job now, not one more tool to help you produce a fourth paper this month, but tools that read the twenty thousand papers you can’t, name the four that matter for your actual question, and refuse to hand you a summary when you ought to read the original. Some of this exists in pieces (including work in my own lab), but almost nobody’s building it as the main event, because “we helped you write less” is a difficult pitch deck.

The human half I would call “slow knowledge,” in the spirit of slow food. It means a few unfashionable commitments. First, adopt a read-to-write ratio and hold to it, because anyone submitting ten papers a year isn’t reading enough to have something to say. Second, give conferences submission budgets rather than unlimited slots. Third, treat synthesis, replication, and reviewing as real scholarly output instead of volunteer labor scraped off the side of a career. Fourth, when you reach for a model, consider asking it to cut your draft in half rather than double it.

Where does the food analogy break down? With calories, you can see the problem: you know what a doughnut is, you can read the label, and the damage eventually shows up on a scale. With information, you can’t tell that something was empty until after you have consumed it, and often not even then, because there’s no label, no scale, and no obvious morning-after. That makes this harder than nutrition, not easier.

Act Two gave everyone a voice. Act Three has handed everyone a megaphone in a room where nobody can hear. I don’t think the answer is to produce less out of guilt, and I’m not going to pretend I have the mechanism worked out. But somebody has to be on the reading side of this, and right now the incentives are paying all of us to stand on the other one. The most radical thing you can do this week is read one thing all the way through, and then tell someone why it mattered.