A proof is either valid or it isn’t. A sentence either moves you or it doesn’t. The provenance of a thing shouldn’t determine its worth — but in practice, it does, almost entirely.
I’ve noticed something odd happening over the past two years, and I suspect you have too. A friend shows you a video, a proof, a piece of writing. It’s good — genuinely good. You say so. Then they mention, almost apologetically, that AI helped make it. And something in the room changes. The praise you just gave suddenly feels retracted, even though nothing about the work itself has changed. Only your knowledge of its origin has.
This is strange if you think about it for more than a few seconds. A proof is either valid or it isn’t. A sentence either moves you or it doesn’t. The provenance of a thing shouldn’t determine its worth — but in practice, it does, almost entirely. We’ve built an elaborate, mostly unconscious accounting system where credit is assigned not to the artifact but to the presumed struggle behind it. Remove the struggle, and the artifact becomes worthless currency, even if it buys the exact same thing.
The Reward Prediction Error
Here’s a fact about brains that most people know abstractly but don’t apply concretely enough: your sense of “how impressive is this” is not computed from the thing itself. It’s computed from the gap between what you expected and what you got. Neuroscientists call the signal that tracks this gap a reward prediction error, and it’s a much better predictor of your emotional response than any property of the stimulus itself.
This explains something that used to confuse me about technological wonder. The first photograph must have been staggering — a frozen moment, forever. The ten-thousandth photograph a person saw was just a photograph. Nothing about photography changed between those two encounters. What changed was the expectation, and expectation is a moving target that recalibrates itself the instant it’s fed new data.
AI compresses this recalibration cycle from decades into weeks. When a model first won gold-medal-level scores on International Mathematical Olympiad problems, it was genuine front-page news — an event people argued about for days. A year later, a comparable result barely trends. Not because the underlying difficulty changed — it didn’t — but because your prior already updated to “computers can do hard math now,” so the actual event carries almost no new information. You are, in a very literal information-theoretic sense, no longer surprised. And surprise, it turns out, was doing most of the work of “impressive” the whole time.
Scarcity Was the Product, Not the Skill
There’s a version of this story people tell that I think is subtly wrong. The story goes: we valued human-made things because they were good, and now that AI makes equally good things, we’re forced to confront that our standards were arbitrary all along. But I don’t think that’s quite it. I think we valued human-made things partly because they were scarce, and we mistook the scarcity for the value.
Consider portraiture. For centuries, a painted likeness of your own face was a luxury reserved for aristocrats, because only a handful of trained hands in any given city could produce one, and each took weeks. Today you can generate a stylized portrait of yourself on Midjourney in ten seconds, indistinguishable in surface quality from what a skilled illustrator might have taken a day to draw. The technical output converges; the social meaning collapses. Scarcity and quality had been so thoroughly confounded throughout history that we never had to distinguish them. Only a rare person could paint well, so paintings were rare, so rare good paintings accumulated enormous social meaning. The rarity was never separately priced — it was baked silently into the admiration. Now the confound breaks apart, and we discover, a little uncomfortably, how much of what felt like “appreciating quality” was actually “appreciating rarity.”
We don’t have cultural machinery built for that. We’re improvising it in real time, badly, mostly by devaluing anything that smells like AI regardless of its actual merit — a crude heuristic that happens to correlate with truth often enough to survive, even though it will produce more and more false negatives as the technology improves.
Tolerance Wants a Bigger Dose — But Whose Dose?
Here’s the part that seems to contradict itself at first glance: today’s mainstream AI output is famously not extreme. It’s sanded down by design — cautious, hedged, allergic to anything that might offend a brand safety filter. So if AI content is mostly bland by construction, where does the appetite for more extreme stimulation actually come from?
The answer, I think, is that AI isn’t generating the extremity directly — it’s generating the flood that makes extremity necessary as a differentiator. When competent-but-median content becomes free and infinite, the only way to still register as a signal is to move to a part of the distribution AI can’t or won’t cheaply occupy. That’s not usually novelty of craft anymore, since craft has been commoditized. It’s intensity: more shocking, more transgressive, more emotionally unfiltered. Human creators, competing for attention against an ocean of AI-generated median content, are pushed toward the tails of the distribution simply because the middle has been paved over. The machine doesn’t need to produce the extreme content itself. It only needs to make everything else feel too cheap to notice.
This is the same curve you see in any system built around variable reward, from slot machines to social media feeds to substance tolerance. The threshold for “interesting” doesn’t reset downward when you get bored of the median — it resets upward, dragging the whole distribution of content that gets attention along with it.
I don’t think this ends in dramatic societal collapse — societies are more elastic than that narrative implies, and there will be a genuine backlash market for scarcity, provenance, and verified human effort, the way there’s now a market for “handmade” and “small-batch” in physical goods that industrialization made abundant. But that backlash market will likely stay a niche for a while, not a norm, and the median experience in the interim is a population whose baseline for stimulation keeps ratcheting up faster than its capacity to actually process what it’s consuming.
That gap between intake and digestion is, I suspect, one of the defining low-grade illnesses of the next decade — quiet enough that nobody names it while it’s happening, obvious enough in retrospect that everyone will wonder why we didn’t. Its symptoms won’t look dramatic at first: a mild, persistent boredom that no amount of new content quite relieves; a fading ability to sit with anything unremarkable long enough to notice its texture; a kind of cultural amnesia in which last month’s astonishment is this month’s background noise. None of that reads as crisis. It reads as normal Tuesday. Which is exactly the problem.