Mona Lisas and Readymades: Provenance is not Authorship

A former colleague messaged me this morning with a question that has become ordinary.

He gets Claude vibes from my LinkedIn comments. Have I automated my socials?

I told him the truth that does not fit a purity test: not fully. I am usually testing a few tools at once for different workflows. Depending on the surface I am on and the caffeine level, your mileage will vary.

He understood. He has gone through his own waves of Claude for communication. Right now he is leaning toward less automated writing, except for CV tailoring and first messages. He is thinking a lot about this.

So am I. Not because the question is rude. Because the question treats surface texture as a verdict on whether anyone is still behind the words.

Claude vibes as evidence

The accusation is a compression. Smooth cadence, tidy structure, a comment that arrives a little too ready: therefore a model wrote it. Therefore the person did not mean it. Therefore the exchange is fake.

Sometimes that compression is right. Plenty of feed sludge is a button press with nobody behind the wheel. Detectors and vibes catch some of that, and they miss plenty else. People have always written bland, over-structured, strangely polished prose without a model in the room. Editors have been deleting insight-shaped filler for decades. The tools did not invent bad writing. They made a certain kind of it cheaper and more common.

What the vibe test cannot settle is the question people actually care about under the word “authenticity.” Did someone exercise judgment over these words? Will they stand behind the claims?

Provenance panic asks how the text was manufactured. Authorship asks whether someone took responsibility for what it claims.

Those are different questions. Provenance is not irrelevant. It is an insufficient proxy for authorship.

The postcard and the mustache

In 1919, Marcel Duchamp took a cheap postcard reproduction of the Mona Lisa, drew a mustache and goatee on it, and titled the result L.H.O.O.Q. Museums later called works like this rectified readymades: manufactured objects selected, altered, framed, and named. He had already shown manufactured objects as art without much alteration at all: a bicycle wheel on a stool, a urinal signed and placed in a gallery.

Nobody looking at L.H.O.O.Q. mistakes it for an untouched Leonardo. That is why the postcard is a poor proof that invisible AI assistance is automatically fine. It is a strong proof of something else. The readymade broke the easy equation between authorship and fabrication. Duchamp did not manufacture the urinal or paint the Mona Lisa. His authorship lived elsewhere: selection, alteration, framing, naming, and the willingness to present the result as his work.

That tradition did not settle every ethical fight about appropriation. Later artists who rephotograph or re-present other people’s images still force hard questions about credit, consent, and extraction. Those fights matter. They do not restore the old equation that authorship equals making every physical or verbal component yourself.

AI tools sit closer to the postcard than to the paintbrush myth. They offer manufactured language. You can paste the postcard and walk away. Or you can decide what it is for, cut what does not belong, put your name under it, and accept the consequences when someone replies.

Traditional authorship culturally bundled three things that generative systems pry apart: manufacture, judgment, and responsibility. The model can increasingly perform manufacture. A human may retain judgment. The named author retains responsibility, or they do not.

That split is what I mean by Mona Lisas and Readymades: authorship as judgment and responsibility over materials that already came from elsewhere. The material might be a model draft, an editor’s pass, a ghostwriter’s outline, or a voice note transcribed at speed. Tools are allowed. Publishing without weighing what you put into the world voids the vow.

Human-generated is not the same as authored. A person can type every word and still abdicate judgment: corporate slurry, received wisdom, claims never examined. A person can use a model and still author the result: selection, transformation, intention, accountability.

The Job Test

When the authenticity panic is loud, start here:

Did it do the job? Did you exercise judgment? Will you stand behind it?

Those three questions are the Job Test. The first covers utility: entertain, inform, persuade, explain, clarify for the job claimed. The second and third cover accountability: do you own the claims, choices, errors, omissions, and the way authorship is represented?

Utility alone cannot carry authorship. A fabricated medical article can educate convincingly. Propaganda can entertain. A ghostwritten confession can move someone while misrepresenting who lived it.

If the piece fails utility, no amount of human typing will save it. If it passes utility while no one will own the claims, the percentage-of-AI score still has not answered the authorship question.

Substack’s detection feature can estimate whether text looks human-written, AI-assisted, or AI-generated. That may be useful evidence about production provenance. It cannot establish intention, comprehension, responsibility, or whether a writer even agrees with what appeared under their name.

Reid Hoffman’s useful line, in the recent Substack fight over those scores, is that a name on a page is a vow. You invested energy. You are present in the words. That vow never required typing every letter by hand. We did not ask that of the novelist who dictates or the leader who works with a ghostwriter. The true failure is hitting a button, declining to weigh or even read the output, and shipping it as yours.

Jeremy Utley, speaking with Steven Johnson on Beyond the Prompt, recalled Guy Kawasaki’s formulation of the same reader contract: readers hire you for the best possible book for their consumption, not for a personal typing quota.

Tools can help manufacture. They cannot accept the vow for you.

Between purity and surrender

Martin Fowler published a short, honest piece this month: he does not like LLMs. The uncanny valley of LLM-voice grates. They bullshit with confidence. He also cites Jessica Kerr’s harder claim: there are situations where failing to use these systems becomes irresponsible because they are both faster and more thorough.

Aesthetic dislike and instrumental judgment are different questions. That combination is more adult than either purity or surrender.

Purity says that if AI touched it, authorship is compromised. Surrender says that output quality is all that matters. The narrower position is to use whatever materials you choose, including less automation on some surfaces the way my former colleague is doing with most job-search writing. Exercise judgment. Accept authorship-level responsibility for the artifact.

Disclosure still matters when policy, trust, or the genre requires it. Fraud still matters. Synthetic people pretending to be human still matter. None of those are solved by vibes. They are solved by judgment, norms, and sometimes labels. Publishing unread output fails the Job Test cold. Provenance estimates remain useful evidence. They remain a poor substitute for the vow.

After Sensemaking Commons, this is the creator-side sequel. Shared referents help a group think together. Authorship is what you owe a single reader when the materials already came from elsewhere.

Claude vibes may tell you something about surface provenance. They cannot tell you whether authorship was actually exercised. The better question is not whether you can smell the machine. It is whether there is still a person behind the words making choices and accepting the consequences.

Madam I’m Adam

This continues the thread from Sensemaking Commons, where shared inputs became a design choice. Here the design choice is authorship: manufacture, judgment, and responsibility when the postcard is already half-made.