Ultra-Processed Insight
What candy did to fruit, AI is doing to ideas.
For most of human history, sweetness was a find. A ripe mango was a pack of calories and vitamins worth the search. The instinct that pulled us toward it was a good instinct. Because, in that world, sweetness and nutrition were the same thing.
Today, they can be pulled apart. The sweetness signal is manufactured, pressed into jelly beans and gummy bears, and sold to a body that still reaches for it based on evolutionary memory. The fruit was rare and valuable. The candy is abundant and almost free. Along the way, the signal came loose from the thing it used to mean.
This is the cycle that runs whenever we get good at producing what we want. We find the source, learn the signal, and figure out how to deliver the signal without the substance. What was rare becomes cheap, what was cheap becomes abundant, and the signal that once pointed at the thing gets manufactured to pass for it.
AI is now doing this to our information diet.
Hostile environments
We see this clearly in food. The mango was sweet for a reason. The sugar came wrapped in fiber and cell walls and a structure the gut had to take apart slowly. You couldn’t eat too much sugar at once because you had to chew through everything else first. In the book, Ultra-Processed People, Chris van Tulleken calls that surrounding structure the matrix. It was the scaffolding that packaged calories.
The processing project of the last century has been the dismantling of that matrix. We learned which signals the body responded to and figured out how to make each one cheaply: sweetness from corn, fat from seeds, mouthfeel from gums and starches. The fibrous structure that used to slow you down got pulverized out. And to keep the label credible, we re-injected the vitamins.
The end state is pre-chewed goop that turned food into a product. It looks right in the photograph and triggers the same instincts the fruit did. With the matrix gone, the body no longer knows when to stop. And the shelf has more of it than any body could finish.
Philosopher C. Thi Nguyen has a name for this. He calls it hostile epistemology: the study of environments that exploit the shortcuts we use. A hostile environment is one that figures out the signal and builds itself around it, so the instinct that evolution built to keep us alive can be coaxed to pull us past where we should stop.
Insight, fortified
Insight used to take work, the way fruit took work. Some of it was the finding. Most of it was the chewing: reading and re-reading, tracing the author’s intent, working out where the idea fit in what you already knew. That work was the matrix. Now it comes pre-chewed, ready to retweet.
The signs that something was worth engaging with used to be attached to the thinking behind it. Credentials meant the author had spent years with the thing. Urgency meant something real was at risk of being missed. A new term meant a new discovery. The industry has coopted all three: credentials into bot-driven likes and retweets, urgency into clickbait (”you’re using Claude wrong”), inventing new scientific-sounding terms.
Take email apnea, a phrase coined to describe the shallow breathing people do while checking their inboxes. The observation may be real. But the phrase borrows weight. Apnea is a clinical term, and attaching it to a mundane behavior gives that behavior a texture of noteworthiness it hasn’t earned. Once you see the move you see it everywhere: a recent piece called AI’s flattening of voice semantic ablation, a fifty-dollar phrase where a fifty-cent one would have done.
What we never saw in the past, in the wild, was this much signal. Writing a credible-sounding headline used to take a human. AI collapsed the cost of all of it. Now there are this many credentials in every byline, this many coined terms, this many headlines warning that we were doing it wrong. We still reach for the signal because that’s what we evolved to do: use it to decide what’s worth our time. But when the signal is everywhere, without the substance, you can spend the day reading signals and never get any value. And the feed, like a bag of chips, never closes.
You finish the essay and you feel fed. A week later you can’t remember why.
The bouncer
Years ago, Nicholas Carr pointed out that the Internet’s real problem isn’t information overload but filter success: the better our filters get at giving us what we want, the more of what we want we get. The needle-in-the-haystack problem became its inversion. We have haystacks of needles now, every niche we’d ever cared about saturated with AI-generated content tuned to look like signal.
But the AI that floods the feed can also stand at the door of your attention. It can act as a bouncer, questioning the credentials, toning down the urgency, tracing the primary source for the claims. I’ve built one of these as a Claude Code skill. It sits on top of my Obsidian vault, the second brain I’ve been writing into for years. When I find an article that looks interesting, I send it through the skill first. It searches the vault for what I’ve already written about the topic, surfaces the contradictions and the confirmations, and tells me whether the new piece is offering something I don’t already have. Most of the time, it isn’t.
I spar with the AI, pushing on the idea, disagreeing, contrasting it with other things I know. Often the verdict is that the idea isn’t worth keeping. Collecting is not knowing. When an idea does survive, it gets connected: joined to what it relates to, contrasting with what it disputes, traceable to its source. The idea compass framework handles the joining.
Signal and substance came apart because we got good at producing signal. AI is the first tool that can also put them back together. The shelf will keep filling. The vault doesn’t have to.