AI Kept Your Meaning and Deleted Your Signature

Read three emails from three different operators this month. Polished, confident, on-brand. Correct verbs, tight structure, not a wasted word. You could not tell them apart if you tried.

That is not a coincidence, and it is not a compliment. There is now a number attached to it.

The study

Researchers led by Sourati and colleagues pulled over 880,000 texts across seven datasets - social media posts, news writing, scientific abstracts - and measured what happened to writing-complexity variance before and after large language models became a standard editing tool. Published in Nature Human Behaviour in 2026, the finding is specific and uncomfortable: since ChatGPT’s release, variance in how people write has dropped 21 to 50 percent, depending on the dataset and model measured.

The mechanism is not “AI writes badly.” It is the opposite, and that is the problem. When researchers ran texts through AI polishing and compared the rewrite to the original, the meaning held: above 95 percent similarity in 87 percent of cases. The content survived. What did not survive, at a measurable rate, was the writer. Models trained to infer an author’s gender, age, ideological leaning or moral values from their writing got roughly six percentage points less accurate once the text had been through an AI pass. The polish amplifies whatever reads as fluent and dominant in the training distribution, and suppresses whatever reads as individual.

Nothing about that process looks like damage. That is exactly why it compounds.

The message survived the edit. You did not. And nothing about the process was designed to tell you that.

This is the law you already know, arriving one generation early

If you read The Universal Law of Copy Degradation, you already have the mathematics for this. Shannon’s Data Processing Inequality states a fact that has no exceptions: no transformation of a signal can create information. Processing can only preserve it or lose it. That piece proved the law holding across three unrelated substrates - Wakayama’s cloned mice failing structurally by generation 58, Shumailov’s AI models collapsing by generation 9 when trained on their own outputs, and a firm’s institutional knowledge going untraceable to any primary source by generation 4 of AI-assisted summarisation.

All three examples were closed loops: a system copying its own prior output, generation after generation, until the tails of the original distribution disappeared for good.

This new study is not a closed loop. It is generation one. A human writes a genuinely original sentence - their own vocabulary, their own rhythm, their own unstated assumptions about who they are talking to - and it goes through a single AI pass before it ships. The law does not wait for nine generations to start operating. It starts on the first transformation. Most professionals who lean on AI for editing are not approaching the collapse point Copy Degradation describes. They are already inside it, at Generation 1, today, in the outbound message they are about to send.

That matters more than it sounds like it should, because Generation 1 is not a milestone on the way to the collapse point. It is the only generation where the primary signal - the version with the highest information content, in Shannon’s exact sense - still exists somewhere for comparison. A cloned mouse at generation 58 has no living generation-zero ancestor left to check against. A collapsed AI model has no access to the original training distribution once the tails have vanished. Your unedited first draft, by contrast, still exists, for a few seconds, in the document history or your own memory of writing it, before you accept the rewrite and it doesn’t. Generation 1 is not early in the collapse. It is the last point where the loss is still reversible, which makes it the only generation where noticing is worth anything at all.

The Amplifier Trap has no opinion about what it is amplifying

The Amplifier Trap names a different failure with the same signature: AI does not fix a bad process, it runs the existing process at machine speed. The amplifier has no opinion about the quality of what it is amplifying. Point it at a broken decision-making habit, get a faster broken habit. Point it at a good one, get a faster good one. The tool is indifferent. That indifference is the entire mechanism - not malice, not a bug, just a system that optimises for the metric it was given and treats everything outside that metric as noise to smooth out.

Writing-assistance AI was given one metric: does this read well and say what you meant. It is very good at that metric. It has no metric for “does this still sound like you,” because nobody asked it to optimise for that, so it optimises it away by default - the same way a compressor optimises away whatever isn’t load-bearing to the signal it was told to preserve.

This is an operator-level instance of the Amplifier Trap, not a new mechanism and not a competing one. The trap’s canonical definition does not change: it is still about AI running an existing configuration faster with no evaluation of whether that configuration deserves to run. What this study adds is the first hard evidence of where that indifference lands when the “existing configuration” is a person’s own voice rather than an organisational process. Six percentage points is not a rounding error. It is the sound of an amplifier doing exactly what it was built to do, on the one input nobody told it to protect.

Why 87 percent looks like good news and is not

Every instinct a competent operator has says 87 percent meaning-preservation is a strong result. In most domains it would be. This is the specific trap Copy Degradation already named: reassurance is the delivery mechanism for undetected loss. A tenth-generation photocopy of a photocopy still looks like a document. A cloned mouse still looks like a mouse for fifty-seven generations before the structural failures compound past the point of recovery. The system tells you it is fine right up until it demonstrably is not, because the thing being measured (does this still say the right thing) is not the thing being lost (does this still sound like the person who said it).

For most writing, in most contexts, that loss is a rounding error nobody will ever notice or need to. This is not an argument against using AI to tighten a paragraph. A status update, a routine confirmation, a scheduling email - none of it needs to carry an individual fingerprint, and treating every AI-assisted edit as a threat would be its own kind of Fear Tax. The reassurance trap only bites where the two questions the study measured actually diverge in stakes: writing where the reader is not just checking the facts, but deciding, consciously or not, whether the specific person behind the facts deserves to be trusted.

The obvious objection, and why it doesn’t survive contact with the data

The first response to this study is usually some version of “people write more like each other for reasons that have nothing to do with AI - platform conventions, editorial house style, the general flattening of internet prose.” That objection predates the study and is worth taking seriously, because it would mean the six-percentage-point figure is coincidence rather than mechanism.

It doesn’t survive contact with the specific measurement that produced the six-percentage-point figure. Population-level drift (“people write more like each other now”) could plausibly be explained by platform convention or editorial style. The author-identifiability figure cannot, because it did not come from comparing this year’s writing to last year’s. It came from comparing the same texts to themselves - the original against its own AI-polished rewrite - and measuring whether a model could still infer the same author’s gender, age, ideology or moral values afterwards. That is a paired, within-text comparison. Platform drift has no mechanism for making an individual text a worse fingerprint of the specific person who wrote it, unless the thing done to that individual text is what did it.

What actually disappears when the amplifier runs

The study’s own language - gender, age, ideology, moral values - names demographic categories because that is what a classifier can measure. In an operator’s actual writing, the same erosion shows up as something more specific and more expensive:

• The blunt phrasing that signals you already survived the objection someone is about to raise. Polished prose smooths the edge that made the point land as earned rather than rehearsed.
• The particular way you frame risk, because you have personally paid for being wrong before and it changed how you write about being wrong now. An averaged version of that sentence reads as competent. It does not read as someone who has actually been there.
• The specific register that tells a reader this operator built the thing, rather than commissioned a fluent summary of it. AI polishing optimises toward the register that scores best on average, which is, by construction, the register of someone who did not build anything in particular.
• The asymmetry in how you phrase a compliment versus a correction. Most people are not evenly polished in both directions. An AI pass tends to even them out, which reads as balanced and reads as generic for the same reason.

None of these show up as a fact a reader could name if you asked them what changed. All of them show up as trust, accumulated across every message a reader has seen from you. Trust built from a hundred messages that each lost a small amount of signal does not announce the loss. It simply stops compounding, and an operator who is not looking for it has no reason to notice the plateau.

Where this actually costs a Golden Prisoner money

An operator who built a $2-5M business did not do it with generic competence. Somewhere in the sequence of decisions that got the business from zero to real revenue, a specific, differentiated read of the market, the team, or the moment mattered more than the average read would have. That specificity is not decoration. It is the asset. It is what a board trusts, what a client pays a premium for, what makes an investor believe this operator sees something the spreadsheet does not.

Writing is where that specificity becomes legible to other people. An email, a LinkedIn post, a board memo, a proposal - each one is a small, repeated demonstration that a distinct person with a distinct read is behind the decision. Run every one of those through an AI polish with no check on what got smoothed out, and the study above says something specific and measurable is happening: the demographic and ideological texture that made the writing recognisably yours, not a competent generic operator’s, is being reduced. Not the argument. Not the facts. The fingerprint.

That is the Amplifier Trap operating exactly where a Golden Prisoner can least afford it: on the one output channel where differentiated judgement is supposed to be visible to the people deciding whether to trust it.

Picture the investor update. An operator drafts three paragraphs on why a quarter came in soft, runs it through an AI pass to tighten the prose, and sends it. The polished version says the same thing the draft said - the AI’s job, by design, is to preserve that. What it does not preserve is the specific, slightly uncomfortable phrasing that made the original read like an operator who has personally sat with a bad number before and knows exactly what it does and doesn’t mean. The rewrite reads as competent risk communication. It does not read as this operator’s risk communication, and an investor who has seen six of these updates over two years is pattern-matching on exactly that difference, whether or not they could name what changed.

The amplifier is not trying to erase you. It has no opinion about you at all. That is the entire problem.

What I have staked fifteen years on, and why it is the same discipline

In 2011, within seven days of the onset of new symptoms, paralysis had taken my legs, descending from the navel. Three days later it started climbing, up from the navel toward my chest, until I was breathing with only the top of my lungs. Expert consensus had a clear prognosis for what came next. I have spent the years since insisting on a different instrument: my own direct, unmediated read of what my body was actually telling me, checked against the averaged consensus rather than replaced by it. Motor control returned to the waist. Deep sensation is still returning. That did not happen because the consensus was wrong about everything. It happened because I refused to let the processed, averaged version of the signal fully substitute for the primary one.

That is the same discipline this study is asking every operator to apply to their own writing. Not “distrust the tool.” Check the primary signal against the processed one before you let the processed one stand in for you. The Amplifier Trap and Copy Degradation both describe what happens when nobody does that check: a faster, smoother, more averaged version of you ships in your name, and nothing about the process announces that it happened.

A thirty-second test before you read further

Open your last five sent emails or LinkedIn messages. For each one, ask a single question: at the point where you clicked “improve this” or accepted a suggested rewrite, do you remember the sentence changing, or do you only remember that it got better? “Better” is the Amplifier Trap’s own report on itself - it is exactly what an indifferent optimiser is built to produce, and it tells you nothing about what got smoothed out to get there. If you cannot recall a single specific word choice you made after the AI pass, on any of the five, you are not editing with AI right now. You are shipping its average of you, and you have been for longer than five messages.

The fix is the same fix, applied to a narrower substrate

Copy Degradation’s countermeasure was never “produce less” or “process less.” It was structural: build a non-negotiable point of external, unmediated primary-source injection into the loop, the way sexual reproduction resets Muller’s Ratchet in biological systems by forcing regular contact with an unprocessed source. For organisational knowledge, that meant direct customer conversations, raw market data, unfiltered observation - inputs that had not already been summarised, averaged, or polished by something else first.

For voice specifically, the same structural principle holds, applied narrower:

• Name the checkpoint. Before a piece of writing that carries your name ships, there is a specific point where a specific word choice is yours, not the tool’s. If you cannot name that point for your last five outbound pieces, you are not editing with AI. You are shipping AI’s average of you.
• Track your own generational depth. Copy Degradation’s diagnostic - Generation 0 is a primary source, Generation 3+ is untraceable to one - applies to your own writing exactly as it applies to a knowledge base. A first draft in your own words, AI-tightened once, is Generation 1. A draft where you accepted a suggested rewrite of a rewrite is further along than you think.
• Reserve the highest-stakes writing for the least processing. The board memo, the founding narrative, the message that has to sound like nobody else could have written it - these are exactly the outputs where six percentage points of demographic and ideological signal is not a rounding error. Polish for grammar. Do not outsource the sentence structure that makes the argument sound like you thought of it, because in the writing that matters most, that is not a stylistic preference. It is the entire credibility mechanism.
• Read the rewrite against the draft before you send, not after. The study measured a population, not your specific message - so the only reliable instrument for whether this particular polish cost you something is a direct comparison, made once, before it ships. That comparison is the external, unmediated check. Skipping it is what lets the amplifier run unexamined.

---

Kasimir Hedstrom | MindMastery sovereigncaptain.com