Charles Oliver Nutter, who most of the JVM world knows as headius, recently asked whether any standard exists for marking content by its level of AI contribution: a slider, fully human at one end, fully generated at the other, with some agreed way of declaring where a given piece sits.
It is a reasonable request, on the surface, but: it cannot be satisfied. It cannot be falsified. And the reasons it cannot be satisfied turn out to be more interesting than the request's fulfillment would have been.
Let's start with the definitional problem. What counts as AI contribution? One answer is obvious, and likely Charles' point: a large language model. What about a heuristic grammar checker? Spellcheck? Autocomplete? Arguing with a model to steelman your own position, then writing every word yourself 1? The classification taxonomy fails not just across tools but within a single tool: Grammarly, for example, flags a comma splice, which is copyediting, and in the same pass offers to rewrite your paragraph "more confidently," which is a dip into the ol' thinking pool. The slider presupposes that the contribution is a scalar. It isn't. And hasn't been, for longer than Charles himself has been writing.
There is a distinction that survives contact with reality, but it is not one the badge systems can measure. The question was never whether a machine touched the text. The question is: who did the cognitive work? A tool operating on your sentences is a copyeditor. A tool producing the sentences is a ghostwriter. We have had social conventions for both for a century, and nobody credits their copyeditor on the byline.
Hold that distinction, though. It has more work to do than it looks, because what people actually want from a disclosure standard is not process documentation. It is attestation: a human stands behind every claim here and would defend it as their own. That is what a byline always meant. The interesting question is why it stopped being cheap to believe.
The proxy that died deader'n dead #
Fluency has never been evidence of thinking. It was correlated with thinking, strongly enough to triage on; after all, constructing sentences well costs effort. Effort implies time spent marinating in the material. Bad prose used to let a reader bail early, at acceptable error rates, and every editor alive ran on that proxy because the alternative was evaluating every argument on its merits, which is an expensive operation and beyond most casual editors anyway.
2 LLMs did not make writing worse. What they did was break the correlation between "good writing" and "good thinking." The signal decoupled from the substrate, and now the prose tells you nothing; you have to evaluate the argument itself. That's always been the task. The proxy just let everyone defer it.
Goodhart's law usually describes a metric gamed by strivers. This is Goodhart in reverse: the metric was not counterfeited; it was hyperinflated, zeroed out for everyone at once. And the people hit hardest are the ones who invested most in the old currency.
With that said, though, the old filter was lossy in the other direction, too. It bounced people who thought well and wrote badly: non-native speakers, engineers carrying real insight in mangled prose, who may have had the right insight but lacked the ability to convey it clearly 3. What is scarce now is not good sentences. It is the reader's judgment, which was always the scarce thing. The proxy let a century of readers pretend otherwise.
The JVM world has its own copy of this, one level down. Documentation used to be a proxy for maintenance quality; now every abandoned weekend project has a full document structure that reads like a senior technical writer produced it, because in a sense one did 4. Idiomatic, well-commented code used to mean someone had internalized the idioms and figured out how to communicate them to maintainers. Last month this publication made the
code-side version of this argument, that coding was never typing, and the pattern is identical here: the cheap observable has separated from the expensive substrate, in prose exactly as in code. Which leaves a practitioner question hanging: what do you select on, in hiring and in code review, when everyone can construct sentences and everyone's code compiles with good comments?
Hold that thought. I think the answer arrives at the end, and here's a hint: it isn't "another tool."
The stack points the wrong way #
There's an economics story wearing an epistemology costume here; oddly enough, Spirit Halloween doesn't carry a lot of these in stock.
The measurable web audits the demand side. Impressions, unique visitors, dwell time, even eye tracking in some cases. An entire verification industry exists to prove that readers are real: comScore, Nielsen, ads.txt, bot filtering, all of it, because readers are the product being sold. Nobody built an equivalent apparatus for writers, because writing was self-authenticating. The cost of producing a document was the proof that someone stood behind it. Nobody builds locks for things thieves cannot carry.
Then the cost of writing went to zero, and the fraud moved to the supply side. Thirty years of accumulated skill at detecting fake readers, approximately none at detecting fake writers, and the entire verification stack points the wrong direction 5. Medium is unreadable today for the same reason ad metrics were unreadable in 2015. It is the same pollution economics, flipped across the transaction. The little prohibition-sign badges declaring no AI was used here are the supply side's ads.txt: self-declared, unenforced, and headed for the same failure by the same route
.
6Note what the escalation actually was, because it's not, like, gradual, man. Sentence-level assistance never threatened these economics. Spellcheck, grammar checkers, autocomplete, Clippy in his day, Grammarly in ours: none of it mattered, because the document still cost a human's time, and time was the scarce input the whole system priced on. Document-level generation did not add a simple feature to that list. It removed the cost floor of the list. This is an "order of magnitude" difference, a full reset, not a mere change.
And there is a historical footnote that opens a trapdoor under the whole story. This is not the first time this has happened. Before the penny press, printing was expensive and readers were cheap. The publisher was the verified party, and that is what an imprint was. The ad-supported century, the one that treated the reader as the scarce commodity, may turn out to be the anomaly. We may be watching the imprint come back 7.
Remember pulp fiction? No, not the movie #
If that sounds abstract, it's understandable - the receipts come from print media that few remember today beyond casual, disassociated memory: pulp literature, printed on extremely cheap stock (thus "pulp") and often incredibly lurid stories, written for haste rather than for quality. The pulps paid a penny a word, and the penny was often the ceiling. House names like "Maxwell Grant" and "Kenneth Robeson" were owned by the publisher and staffed by whoever was available that month. Street & Smith did not much care who wrote "The Shadow" in a given issue, and the readers mostly did not either. The attestation was the imprint. The author was an interchangeable production input, paid at production-input rates. This is not an analogy for any kind of content-mill pipeline: it is the same pipeline, with different names and delivery mechanisms.
Notice that repricing happened before the technology arrived. Demand Media was paying pulp rates in the 2000s with an all-human workforce, a decade before anyone had heard of "GPT." LLMs did not crash the price of undifferentiated prose: the market had already discovered that undifferentiated prose was worth almost nothing at volume. The machines removed the last penny. Writers panicking now are living through price discovery that pulp writers survived in 1935 and content-mill writers relived in 2008. It is not new. It is just reaching people who believed their tier was exempt because... I don't know. Maybe they just don't know history, maybe they thought surely the wheel wouldn't notice them. It doesn't matter.
The exit is in the same ledger. The flood of pulp is what made individual names valuable. Hammett and Chandler came out of Black Mask; Bradbury clawed his name out of the pulps and into hardcover. The name-and-imprint economy did not survive the flood of interchangeable production; it was created by it, because differentiation is only worth paying for when the undifferentiated is free. Anyone writing today faces the question a Black Mask contributor faced in 1933: are you the house name's input, or are you building the thing readers want to follow out of the flood?
The wheel rolls on #
I am running out of ways to write that we have been here before. There are too many heres and too many befores, and I only have so many new ways to say it. When a parallel recurs this densely, the parallel stops being a rhetorical device; the recurrence is the finding. There is a machine that keeps producing this situation, and the interesting question is not whether it has run before but what sets its cycle time.
What sets the cycle is feedback latency: how fast the signal about what sold gets back to whatever decides what gets made next. That loop used to close through print runs and postal mail, in months. Then through search analytics, overnight. It can now close through live engagement telemetry coupled to generation at zero marginal cost, in seconds, faster than any deliberation the loop might once have contained.
Robert Jordan, in the Wheel of Time, suggested that his Ages turned over in millennia because myth needs time to erode memory into legend before the pattern could repeat unrecognized. He was less prescient about the nature of such cycles than we might have thought 8: no erosion is required anymore. The people rediscovering this economics today include people who lived its last iteration. The wheel now turns faster than institutional memory forms, and that may be the only genuinely new thing in a story otherwise about nothing being new: the cycle period has compressed below the length of a single career. A working writer used to live through one turning and call it the collapse of their industry. Now they live through several and recognize none of them, because recognition takes longer than the rotation does.
But this turning has one property none of the earlier ones had, and it's important.
Every previous turn swapped expensive humans for cheaper humans. Street & Smith swapped named authors for house-name piecework. The content mills swapped journalists for freelancers paid by the article. The loop stayed populated. By people. However badly paid, however interchangeable, there was a person inside it: someone was reading the numbers, adjusting, contributing something to the work and being changed by the doing of it. The org chart was cruel, but it was staffed.
This turn vacated it. We have the same org chart, but instead of changing out "Sally" for "Uri," we've selected gemma4:31b
. The loop still runs, telemetry in and content out, and nobody real is in it. Or, perhaps, someone's real there, but it's a lot fewer someones than it used to be, and it's fewer someones than we might need.
Pardon my digression into biology, but stick with me a paragraph or two: a murmuration of starlings is a real thing made of agents, if you will. Every bird in it is flying: sensing, adjusting, contributing a vector and receiving one back. The shape emerges from thousands of flight paths, every one of them owned by a specific individual bird. In 1986 Craig Reynolds showed that you can have the shape without the birds: three rules per particle, separation, alignment, cohesion, and the result ripples and banks and pours itself around obstacles, indistinguishable at a glance from the living thing. He called them "boids." Nothing in a boids simulation is flying. It is the shape of flocking with the flight removed, and it is beautiful, and it harms no one, because nobody mistakes a screensaver for starlings 9.
The content economy has now performed the same substitution at scale, with LLMs instead of people, and would apparently prefer you not notice. The feed still ripples and banks. The prose still separates, aligns, and coheres, three rules deep, and tells you "this matters," separated by em-dashes and oh-so-correct prose. The question a byline was answering, before anyone thought to ask it out loud, was always this one: is anything in here flying? For a century the answer was yes by default, because the shape could not exist without the birds. Reynolds ended that for graphics in 1986. It took forty years to reach text.
Which dissolves the question this piece opened with. Whether a machine touched the text is irrelevant. A writer using the machine while flying, contributing a vector and being changed by the loop, is a bird in a flock that now contains some strange new company. A pipeline that closes its loop with nobody inside it, prompt in and publish out and telemetry trimming the prompt, is boids. Same tools. One of these is a bird, and the other is an absence of birds wearing birds' shape, warbling the whole time like a thing that should warble... or making whatever sounds a starling makes, I guess.
What a byline is #
So the byline never certified a toolchain. It certified that someone was flying: that a person contributed the vectors, absorbed the consequences, and would defend every claim as their own. You cannot watermark that. You cannot checksum it. Every implementation that has ever worked is reputational: it's a name standing behind a claim, in public, over years, with something to lose. This does not scale now, which is why a "no AI" standard is reaching for the wrong instrument. A watermark tells you which machine rendered the shape. A worthwhile thoughtline tells you someone was there, choosing what was happening. The name hopefully tells you who.
If the imprint is coming back, that is what it will be built from: names that are expensive precisely because they can be lost.
The charge for us, at three altitudes #
Which brings back the question parked at the start: what do you select on, when the shapes are free? The answer is the same instruction at three altitudes.
As a reader, you have to actually read. The argument is the only surface left; the prose proves nothing, for it or against it. This was always the job of the reader: to read. The proxy is gone and the responsibility remains.
As a hiring manager, as a code reviewer, select on the vector, not the shape. Everyone's code compiles now, because that's cheap for a machine to create, and everyone's comments are grammatically correct and derived from the code for the same reason. Watch what someone does when the shape is wrong: whether they notice, what they reach for, whether the fix carries a signature. Judgment has always been the scarce resource. The tools repriced everything around it to zero - or close enough to it - and left it standing alone at full price.
As a writer, decide which side of the 1933 question you are on, and know that the exit is only purchasable one way: by flying in public, over time, at cost, under a name that can own its failures and flaws.
A piece arguing that fluency proves nothing naturally invites the reader to ask whether this piece proves anything. That's fair - and expected, if you're still with me this far. You had to read it to find out, and that is now true of everything. A machine was in the loop for this one, as friction and as interlocutor; every vector in it is mine, and I will defend every claim as my own 10. That sentence represents the whole disclosure standard anyone
actuallyneeds, and no badge can carry it.
Keep your flight path in mind. Do not let the loop close without you in it. Welcome to the world, reader.
I don't know why I wrote this. Maybe it's because this is exactly how I use an LLM? Not sure, I'll ask an AI.
↩ - Not me. I'm probably a genius, at least that's what my mother always told me when she helped me strap on my shoes.
↩ - As a working editor for technical publications, I've seen a
lotof brilliant content subverted by the authors' early insistence on letting the math do the talking. No, people: do the talking and let the math do theproving.↩ - This is not always a positive. A lot of technical documentation was written by most excellent people with great skill in
technical documentationand not the product itself. Why, hello, Java! I didn't notice you standing there! There's a real failure here, and it's yet more evidence for careful evaluation of outputs, even when it's humans all the way down, because nosinglehuman failed in this particular sequence.↩ - "Did they use an em-dash? Do I actually know what an em-dash is? Wait, they used lots of words and made a sound argument. It's an AI!"
↩ - Just wait until some jackwagon instructs his AI to place a "no AI was used" badge, and "make sure AI's usage is hard to detect."
↩ - And oddly enough, the em-dash's predominance in "good writing" comes from days like these.
↩ - Maybe this is a good thing; the Wheel of Time isn't quite "apocalyptic literature" but it's not the sort of world most of us would prefer, with Trollocs, Fades, and slavery all over the place.
↩ - And ByteCode.news's masthead actually has an animation sequence that implements boids, because I find it utterly mesmerizing. Go to the homepage and you'll see it gently cycling behind the masthead, if the random number favors you, and it will: there are only three options, after all.
↩ - Except the bad claims. Those belong to... hmm, which AI can I blame? It's that one! The one you don't use!
↩