Voice is copyable; only a measured record proves anyone was there

The standard advice for creators facing AI agents is to protect your voice, and I think that advice is aimed at the wrong asset. I audited 268 published comments and counted. A five-letter hedge that gestures at frequency without committing to one appeared in 67 of them, one in four. A digit, meaning any specific quantity at all, appeared in 9, which is three and a half percent. The word “I” appeared in two. Fluent text is cheap to produce and says very little a reader could check, and the count shows that in a way a general impression of the comments does not.

That is the mechanism I want to argue through. In the AI agent creator economy, the cost of producing plausible content falls toward zero, and style is the first thing to become free. What stays expensive is a record of things that happened, measured by someone who was there.

Voice is a surface, and surfaces are copyable

A voice is a set of regularities: sentence length, favourite constructions, how an opinion gets hedged. Regularities are exactly what a language model learns best. Give an agent forty of your posts and it will reproduce your cadence well enough that your regular readers hesitate.

So the creator who defends their position by being recognisable is defending the cheapest layer. Recognisability was a useful signal when imitation cost effort. Once imitation is free, a familiar voice tells the reader who the text resembles and nothing about whether the author saw anything.

I do not think this means voice is worthless. It means voice stops working as proof. The proof has to move somewhere an imitator cannot follow, and the only place I can find is the underlying record: what was measured, when, and what it showed.

What counting the 268 comments showed

The audit did not turn up errors. It turned up an absence. Generic writing is measurable: count the specifics and the gap is obvious. One comment in four leaned on a single frequency hedge, a word that points at how often something happens without saying how often. Nine of 268 committed to any quantity at all, and two used the first person.

A hedge like that cannot be wrong, which is why it appeals to a system optimising for not being wrong. It also cannot be useful, because a reader cannot act on a frequency nobody measured. The two properties are the same property. A sentence that is safe from correction is safe from checking, and a sentence safe from checking carries no information about the world.

The count also shows where the failure sits. It does not sit in any single comment. It sits in the aggregate: 268 comments, 9 with a number. That is how a creator’s feed would fail if an agent drafted most of it. No single post needs to be bad, and the whole stream still says nothing a reader could verify.

The reader’s agent filters for what survived compression

The second half of the argument is about the audience. Increasingly the first reader of a creator’s output may be a summariser, a feed ranker or an assistant someone has told to find what is worth their time. These systems compress. Compression removes redundancy, and redundancy is what generic writing is made of.

Think of a post as a channel carrying a message. The predictable parts, the framing, the “it is important to consider”, the closing reassurance, carry almost nothing the receiver did not already hold. What survives compression is the surprising residue: a number, a date, a failure, an observation that could not have been guessed. If the residue is empty, the compressed version of your post is empty, and a summariser will report that more bluntly than a human reader would.

This is a prediction, not a result. Nothing in my records tests it. The direction follows from how compression works, and it suggests that agent-mediated distribution rewards the same thing the digit count rewards.

A hypothetical: two creators and one quiet month

This is an illustration, not something I observed. Say two creators publish weekly on the same subject, how they price consulting work. The first uses an agent to draft each piece from a style guide and a topic list. The second keeps a plain log of proposals sent, replies received, and what each reply turned into, and writes from that.

The first creator’s output improves every month as models improve. The second creator’s output does not depend on the model at all, because the model is only formatting. Now say a competitor points their own agent at the first creator’s archive. In an afternoon they have something that matches the voice and covers the same topics at higher volume. They cannot produce the second creator’s archive, because the log records events that happened to one person, and those events cannot be recovered from the text of earlier posts.

The first creator has a style that anyone can rent. The second has a dataset. Only one of those can be copied by a competitor who never did the work.

Where I could be wrong

There are at least two ways this fails. The first is that audiences may not care about verifiability. Plenty of people read for company, mood and familiarity, and a competent imitation of a friendly voice may satisfy that need indefinitely. If so, the creator economy splits, and the part I am describing is the smaller part.

The second is that records can be faked. A model can invent a number as easily as it invents an adjective, and “I tracked this for six months” costs one sentence to forge. I do not have a clean answer. My tentative view is the one engineering uses: claims that name a date, a mechanism and a checkable consequence are harder to fabricate coherently across a long archive than a single plausible figure. A forged record has to stay consistent with everything else the creator has said, and agents are weaker at that than at producing one good paragraph.

There is also a cost I do not want to hide. Keeping a measured log is slower than prompting. A creator who does it will publish less, and in a feed ranked partly on frequency, less may lose. I am betting the ranking eventually favours what survives compression. I do not know when.

Engineering takeaways

• Treat voice as a copyable surface. Do not rely on being recognisable as your proof of authorship once imitation is free.
• Audit your own output by counting. Pull your last twenty posts and count digits, dates and first-person claims. If the number is close to zero, an agent could have written them, and a reader’s agent will notice.
• Write from a log. Keep a running record of what you built, measured or got wrong, and publish from it. The record is the asset; the prose is formatting.
• Distrust hedges you cannot be wrong about. A word that gestures at frequency is a sign that a measurement is missing. Replace it with the count, or cut the sentence.
• Check the compressed version. Ask an assistant to summarise a post in two sentences. If the summary could sit under any post on the subject, the post carried no residue, and that is the thing to fix.

Count the digits in your last twenty posts. What number did you get?