The starting point for this piece was modest: I wanted the articles AI helped me write to carry less of that heavy "AI smell."
I have been writing with AI for a while now. The bulk is generated by an agent; I set the direction, review the content, and rework the phrasing. But every time I publish, readers can sense a machine flavor—sentences too even, arguments too clean, word choice circling among a few high-frequency options. So I put real effort into studying what the AI smell is and how to reduce it. I dug through a pile of papers, updated my writing skill, built a vocabulary list and a sentence-pattern checklist. I figured I had the problem nailed.
But somewhere in the research and debugging, some questions I hadn't planned to think about started surfacing. The AI smell makes me uncomfortable—is that because it "feels like a machine," or because of something else? If the AI smell vanished completely one day, would my trust in a piece of writing automatically return? The more I thought about it, the more I suspected that my original question—how to make writing not look like AI wrote it—was aimed in the wrong direction.
The thoughts below began with that tangent.
There is one observation that strikes me as more and more important the longer I sit with it.
Before AI, how polished a product was correlated strongly with its quality. You'd see a beautiful app and instinctively assume the team had put thought into it, and you'd have more confidence in its features and stability. This heuristic isn't always right, but it's right most of the time—making something refined takes time and taste, and people willing to invest there usually build the features well too.
Writing works the same way. An article that is well-structured, clearly argued, and precisely worded was probably written with care. Read two paragraphs and you could make a basic call: this person thought, spent time, is worth reading on.
AI broke that heuristic.
"Looks good" used to be a cheap filtering tool. You'd scan something, eliminate the ones lacking polish first, and save your attention for the rest. But AI raised the floor of polish to a height that costs almost no human time. You hit enter, and out comes an article with complete structure, coherent logic, no obvious word-choice errors. You hit enter again, and out comes an app interface that looks like it went through several rounds of design review. The things that lacked polish have nearly vanished from the information space; everything now sits in a "looks fine" band. The floor has been lifted so high that the information "looks good" carries is close to zero.
At that point, if you want to sort out what's truly worth your time, you have to read. And reading takes three things: time, attention, and judgment.
It's precisely these three things that are being systematically drained.
The total volume of content is exploding. AI has multiplied by several orders of magnitude how much text one person can produce in a day, and most of that text enters the public information space. Where you might once have faced ten new articles a day, it could now be fifty, a hundred. You can't treat each one with the same care—you simply can't. The time left for each piece keeps shrinking.
And judgment itself is degrading. In early 2026, BCG and HBR surveyed 1,488 U.S. employees; 14% reported "AI brain fry"—mental fatigue from continuously supervising and reviewing AI output, beyond cognitive load. In engineering it was 18%. These people had 33% higher decision fatigue, were 39% more likely to make major errors, and were 39% more likely to want to quit. This fatigue is a different beast from traditional burnout—using AI to replace repetitive labor cut burnout by 15%, while using AI for supervision and review raised fatigue by 12%. The drain comes from watching, not from using AI.
An experiment by Zhu et al. in 2024 found that when sources aren't labeled, human accuracy at distinguishing AI text from human text is close to a coin flip. A Nature study the same year was more unsettling: the more people scrolled on phones and social media, the worse they got at telling the difference. Exposure doesn't produce training; it produces numbness. A review by Sourati et al. published in Trends in Cognitive Sciences in 2025 confirmed that stylistic diversity—on Reddit, in scientific papers, in journals—is already measurably declining. AI text, as ambient noise, is quietly reshaping how humans express themselves.
So it isn't that any one thing has gone wrong. Three curves are moving in unfavorable directions at once: the things needing discernment are increasing, the time available to discern is decreasing, and the ability to discern is itself degrading. Three curves—two rising, one falling.
This predicament isn't entirely new.
Every leap in information technology in history has destroyed some old quality signal. In the age of manuscripts, the mere existence of a book was a signal—copying a Bible cost a monk a year of his life. The printing press killed that signal: books were no longer rare, and "this is a book" no longer carried quality information. People turned instead to publisher reputation, author fame, and peer review to filter. The telegraph and telephone meant transmission speed was no longer a signal either—being well-informed once required connections and resources; later a telegram settled it. The internet drove distribution cost to zero: being able to publish a book or get into a newspaper had itself been a signal, and then anyone could start a blog, killing one more signal.
What AI is doing belongs to the same sequence: it has knocked out "cost of production" as a signal. Where "this is written pretty well" used to mean the author put in effort, now it means nothing. It's just the default level that comes with AI output.
Each time an old signal failed, humans went through a period of chaos—and then gradually rebuilt new filtering mechanisms. Publisher reputation took roughly a century to stabilize. Peer review took more than a hundred years to go from seed to academic standard. The social filtering of the internet era—looking at who reshared, the quality of the comments, the author's track record—is still being refined today. We're probably in the chaotic phase of the AI signal collapse, still a long way from new mechanisms settling in.
Seen from this angle, the earlier intuition about the AI smell was right, but the reason needs correcting.
We resist the AI smell not because it comes from AI. If AI only handled phrasing and structure while the content was your own thinking, you wouldn't resist it so much. You resist because the AI smell builds a wall between you and the content—it tells you "this writing has been processed by AI at the level of expression," and then nothing more. It can't answer the question you actually want answered: is this worth reading? The AI smell carries zero information.
This also explains why we miss the small flaws in human writing. Those awkward phrasings, abrupt turns, sudden shifts in paragraph rhythm—they aren't aesthetic virtues. They used to be a cost signal: a flaw proved the author made a choice, the choice proved the author put in thought, and thought meant there was something here worth your attention. A polished sentence a human writer arrived at through repeated revision and a concise sentence an AI was statistically bound to produce can look identical and carry completely different information. The former tells you "this person thought"; the latter tells you "the model computed." But you can't see the difference. The signal isn't open to you.
Back to ourselves.
I don't think everyone needs to fight this trend. You can care or not care, depending on what you value. But if you want to keep your judgment about information quality rather than being passively washed along by the stream, here are a few things I'm trying.
The first is to unbind "polish" from "quality" in my head. This is harder than it sounds. AI-generated things look so complete that the brain automatically treats completeness as a proxy for quality. When I read an article, a PR, a design doc now, I deliberately ask myself a dumb question: if its expression were polished to perfection, what would be left of the content itself? If the answer is "not much," then no matter how beautiful the expression, it's only expression.
The second is to look at choices, not at correctness. The defining feature of AI output is that every sentence is correct, but not one had to be said that way. Human writing is the opposite—there will be flaws, there will be inaccurate calls, but certain phrasings are ones the author had to write that way, because that was their judgment, the thing they truly wanted to stress, the place they wouldn't compromise. I now work to find those moments of "it could have been said otherwise, but they insisted on this." In an era when anyone can produce correct expression with AI's help, choice is the signal.
The third is the most basic: control your pace. You don't need to invest equal judgment in every piece. In fact, you can't. For most things, a quick scan is enough. Save full judgment for the people you're willing to bet on—not because of their title, but because what they've written before made you believe they're worth the time.
The article I'm writing right now is caught in the same bind.
If I don't tell you which parts are my thinking and which the AI helped me organize, you have no way to know. The opening observation about completeness as a signal formed slowly as I debugged my agent writing skill. The research data in the middle was found by AI. The conclusions at the end I reasoned out myself. But whether what I'm telling you is true, you can't verify.
So whether this article is worth reading comes down, in the end, to whether you're willing to trust me.
In a world where content can be produced without limit, the truly scarce things won't accelerate just because the technology does. Judgment remains scarce. Trust remains scarce. A person willing to spend time thinking rather than time generating remains scarce. Possibly more scarce than before AI arrived.
A few things I still haven't figured out
After the explosion in total volume, what is the ratio of good content to bad? My intuition is that both are growing, but not proportionally—bad content grows far faster than good, because good content needs judgment for the final filtering step, and judgment is being drained. I have no data on this. Does the ratio have an equilibrium point of its own, or does it keep deteriorating? If it trends toward zero, with good content becoming a mathematically arbitrarily small fraction, then "looking for good content" becomes irrational in itself—no different from searching the ocean for one specific needle.
From hand-copying to the printing press, to the telegraph and telephone, to the internet: each time a quality signal was destroyed, how long did it take humans to rebuild filtering mechanisms? Publisher reputation, roughly a century; peer review, more than a hundred years; social filtering, still in progress. What's the answer now? Substack-style personal reputation? Something that hasn't appeared yet? In the article I mentioned "looking at the author's track record," but honestly that's too slow, and it only works for people who keep producing. How does a newcomer build trust? How does someone who writes occasionally but has real substance get discovered? I have no answers to any of these.
There's also a question of timescale. I said just now that the AI smell carries zero information—but if future human writing is assimilated by long-term exposure to AI, might the AI smell instead become the default feature of "human writing"? What signal would distinguish humans from AI then? Writing with flaws, with looseness, with personal verbal tics—might it instead be optimized away by algorithms, because it's statistically deemed "not good enough expression"? This article may be unreadable in three years—not because the conclusions are wrong, but because the things these words point to—"completeness," "quality," "AI smell"—may by then no longer be what I'm discussing today.
References
- Bedard, J. et al. (2026). AI Brain Fry: Managing Cognitive Overload in the Age of Artificial Intelligence. Harvard Business Review. Survey of 1,488 U.S. employees; 14% reported AI brain fry, 18% in engineering, associated with 33% higher decision fatigue, 39% higher error rates, and 39% higher intent to quit.
- Zhu, T., Weissburg, I., Zhang, K., & Wang, W. Y. (2024). Human Bias in the Face of AI: Examining Human Judgment Against Text Labeled as AI Generated. arXiv:2410.03723. Detection accuracy near chance without labels; labels can manipulate judgment.
- Nature Scientific Reports (2024). Human intelligence can safeguard against artificial intelligence. Heavy social media and smartphone users are worse at distinguishing AI text; exposure produces habituation rather than training.
- Sourati, Z., Ziabari, A. S., & Dehghani, M. (2025). The Homogenizing Effect of Large Language Models on Human Expression. Trends in Cognitive Sciences. Stylistic diversity is measurably declining on Reddit, in scientific writing, and in journals, with a recursively structured feedback loop.