I keep seeing a strange new genre of criticism.
Someone posts an essay. A reader notices an em dash, a suspiciously balanced paragraph, perhaps the word “delve,” and announces:
I can tell this was written by ChatGPT.
This is harmless literary sport until the same instinct migrates into science, universities, journalism, and other places where we supposedly care about whether something is true.
Was this paper written by AI?
Did the scientist use Claude?
Did the student really type those sentences?
These questions sometimes matter. If an assignment tests unaided writing, the rules matter. Fabrication, plagiarism, false claims of authorship, and confidentiality violations matter. The same anxiety is now showing up around AI-generated conference and journal reviews: if a reviewer delegates the judgment to a model, what exactly is the reviewer contributing?
I see where the concern comes from. Peer review depends on expertise, attention, and accountability, and nobody wants those quietly replaced by fluent autocomplete. But we are allowing the provenance question to swallow a more important one.
Is there anything good underneath the prose?
Is the claim true? Is the evidence good? Does the conclusion follow? Is there actually a new idea here?
Generative AI is making competent-looking symbolic production extraordinarily cheap.
Our response should not be to spend the next generation reconstructing the old scarcity by determining which sentences were handcrafted.
We should figure out what was valuable underneath them.
AI did not invent slop
There is unquestionably an AI-slop problem.
It is now possible to manufacture enormous quantities of plausible-looking text for almost nothing, and we are seeing the predictable consequences. Recent analyses have found large numbers of apparently nonexistent citations in scientific papers. Courts are dealing with filings containing invented cases and other AI-generated material.
This is bad.
But notice what is bad about it.
The citation does not exist. The case does not exist. The reasoning is wrong. The author did not check the work.
Those are substantive failures.
“AI generated it” explains how those failures became extraordinarily cheap to produce. It is not itself the failure.
Humans have produced vacuous papers, terrible arguments, fabricated evidence, corporate mush, and confident nonsense for a very long time.
AI did not invent intellectual slop.
It industrialized its production.
That distinction tells us what problem we actually need to solve.
We need to become better at detecting slop.
Not better at detecting AI.
Science already knows how to do this
Suppose I give you a mathematical theorem and a proof.
You do not know whether the proof was discovered by a mathematician, suggested by an AI, or produced through some collaboration between the two.
What should you do?
Check the proof.
If an inference is invalid, the proof fails. If every step works, the theorem follows.
There may be separate questions about attribution, originality, disclosure, or competition rules. Those can matter enormously.
But they are different questions.
The same principle applies to science.
Did the experiment happen? Is the data real? Does the analysis support the claim? Can somebody reproduce the result? Does the cited paper actually support the sentence?
These are scientific questions.
Modern science was constructed partly around the recognition that authors cannot simply be trusted. A proposition does not become true because a famous scientist says it. An experiment does not become valid because it came from an elite university.
That is why we have proof, experiment, reproducibility, citation, peer review, measurement, and adversarial criticism. I have made the same argument in AI security: start with concrete threat models and testable failure conditions, not slogans.
We built an intellectual culture around a wonderfully disrespectful principle:
Show me.
And now, having constructed all that machinery for evaluating claims independently of their authors, we seem tempted to respond to generative AI by asking who typed the commas.
That is backwards.
The real thing AI is destroying
For most of modern history, polished intellectual artifacts were expensive.
Writing twenty coherent pages took time. Producing a competent scientific paper usually implied that somebody had read things, thought about a problem, organized an argument, and revised it.
Fluent prose therefore carried information beyond its literal content.
It was a weak proxy for effort, education, expertise, and thought.
Generative AI breaks that correlation.
A machine can manufacture the surface appearance of intellectual labor without necessarily performing the intellectual labor we imagined underneath it.
That is genuinely disruptive.
But the mistake is to conclude that we therefore need to restore the old signal.
We cannot.
The signal is becoming cheap.
Versailles
There is nothing particularly new about this process.
When Louis XIV built the Hall of Mirrors at Versailles, enormous high-quality mirrors were objects of extraordinary expense and technical achievement. Part of their magnificence came from the simple fact that making them was extremely difficult.
Today an ordinary middle-class household can contain mirrors everywhere.
The mirrors did not become less real.
What disappeared was their scarcity.
And with that scarcity disappeared much of the status attached merely to possessing one.
Technology repeatedly does this. It takes something whose difficulty of production has become entangled with its value and makes the production cheap.
The dangerous moment comes afterward, when we confuse disappearing scarcity with disappearing value.
Generative AI is doing this to fluent symbolic output.
For centuries, polished prose was difficult enough to produce that the artifact itself provided weak evidence that substantial effort had occurred behind it.
That evidence is disappearing.
We can spend enormous energy sorting sentences into “authentic” and “synthetic.”
Or ask the more interesting question:
Once the surface becomes cheap, where does the value move?
For science, the answer is clear: truth, novelty, evidence, reasoning, prediction, explanation, judgment.
Those were the valuable things all along.
“Human-made” is not an argument
I do not think creative work gets a permanent exemption either.
People are perfectly entitled to value provenance. Someone can prefer handmade furniture or value a poem partly because a particular person wrote it.
But that is different from claiming that an artifact is better merely because a human produced it.
“Human-made” may be a preference. It is not an argument.
If human writers produce something machines cannot, show me the difference. If lived experience, intention, taste, embodiment, or consciousness creates something distinctive, wonderful. Identify it.
Humans do not get to win by definition.
There is always another fence: machines can calculate but not reason; reason but not play chess creatively; play chess but not write; write but surely not produce real literature.
Some of those boundaries may capture genuine differences.
If so, show the difference.
Simply declaring a protected category called “human creativity” does not settle the question.
There is also a technical problem with AI detection
Even if we wanted to reconstruct the old boundary, there is a deeper reason to doubt that stylistic AI detection is a durable solution.
An AI-writing detector does not detect some metaphysical substance called “machine authorship.”
It detects differences.
Perhaps current models use certain phrases too often. Perhaps their sentence lengths have a recognizable distribution. Perhaps they prefer particular transitions, structures, punctuation marks, or rhythms.
A detector can exploit those regularities.
But now suppose people start using the detector.
The obvious response is to change the generator.
If em dashes trigger the detector, stop using em dashes.
If paragraphs are too balanced, vary them.
If particular phrases are suspicious, remove them.
If the detector itself is available, generate ten versions and choose the one it considers most human.
Then the detector adapts.
Then the generator adapts again.
This is not a static classification problem.
It is an adversarial game.
And the generator has an enormous advantage: its entire objective can be changed to produce text increasingly indistinguishable from whatever humans currently produce.
The very success of a detector creates pressure to erase the feature it detects.
Yesterday’s AI tell becomes tomorrow’s model-training target.
Meanwhile humans are moving too. People adopt expressions they encounter from models and consciously avoid phrases associated with ChatGPT. Human and machine styles influence one another.
So the target moves from both directions.
There is no reason to expect a permanent punctuation pattern, vocabulary distribution, or prose rhythm that says: machine.
At some point, if the outputs become sufficiently similar, the text alone simply does not contain enough information to reliably reconstruct how it was produced.
That is not a failure to invent a sufficiently clever detector.
It is a failure to have enough evidence.
If provenance matters, use provenance
Sometimes provenance really is what we care about.
Did the researcher perform the experiment? Did the journalist conduct the interview? Was the student required to work unaided? Was the image captured at the scene?
Fine.
Then establish provenance.
Use disclosure, records, authentication, chain of custody—whatever the domain requires.
Do not pretend prose style is a chain of custody.
And do not confuse provenance with quality.
A perfectly authenticated human-written paper can still be nonsense.
An AI-generated argument can still be correct.
These are different axes.
Accountability is more interesting than authorship
There is a better question than:
Did you personally type every word?
Ask:
Will you stand behind every claim?
If an AI puts a fake citation into my paper, that is my problem. If it changes the meaning of a theorem and I fail to notice, that is my problem. If it gives me an elegant argument with a logical hole and I publish it, that is my problem.
And if I put my name on a paragraph I cannot explain, that is embarrassing whether it came from ChatGPT, a research assistant, a coauthor, or my own sleep-deprived typing.
Accountability tracks something important.
Token provenance often does not.
AI should therefore make science stricter in the right direction.
If papers become cheaper to generate, make them easier to interrogate.
Show the data. Show the analysis. Show the proof. Show the source. Make the result reproducible. Make the author defend it.
The answer to cheap prose should be stronger verification, not more elaborate authorship divination. The same distinction matters in security: a mechanism is not a guarantee merely because it has the right label—a kill switch, for example, is not the same thing as erasing the information paths that let a system reconstruct its prior state.
The real slop detector
So yes, worry about AI slop.
I do.
We are going to live in a world containing quantities of plausible-looking material that would once have been economically impossible to manufacture.
But:
Slop is a quality category, not a provenance category.
A paper containing twenty polished pages and no new idea is slop. An argument whose conclusion does not follow is slop. A citation that does not support the sentence is slop.
And a beautifully written human essay that says nothing does not become valuable because a human suffered while typing it.
The durable detector asks harder questions.
What exactly is the claim?
What is new?
What evidence supports it?
Does the conclusion follow?
Can the result be reproduced?
Can the person whose name appears on it explain and defend it?
Those questions are harder than counting em dashes.
That is the point.
The Socratic test
Socrates did not have an AI detector.
He had something better.
Someone made a claim. He asked what it meant, why they believed it, what followed from it, and where it contradicted itself.
The identity of the speaker was secondary.
A human says something.
Interrogate the claim.
An AI says something.
Interrogate the claim.
A scientist and an AI produce something together.
Interrogate the claim.
The machinery of intellectual production is changing dramatically.
The epistemic standard does not have to.
The age of AI is the end of being able to use the cost of producing an intellectual artifact as a reliable proxy for its value.
For a long time, fluent language, polished presentation, and sheer quantity of work provided weak evidence that somebody had thought hard.
Machines are destroying that shortcut.
Good.
Perhaps we relied on it too much.
The durable questions were always the older ones:
Is the claim true?
Is the evidence good?
Does the conclusion follow?
AI did not make those questions obsolete.
It made our excuses for not asking them obsolete.
Author’s note: I use AI tools for editing and research assistance; the arguments and views here are my own.