George Orwell wrote three documents about controlled language and almost everyone remembers only two of them. From 1942 to 1944 he was at the BBC promoting Basic English, C. K. Ogden's 850-word stripped-down vocabulary, as a bridge for listeners across the Eastern Service. In 1949 he published the appendix to Nineteen Eighty-Four, in which a language with the range engineered out of it engineers the range out of the people using it. Those are the two everybody cites, usually to argue that Orwell changed his mind.

The document in between is the one that actually resolves it. In 1946 he published "Politics and the English Language," which ends with six rules for writing clearly. Five of them are prohibitions. The sixth says to break any of the other five rather than say something outright barbarous.

That sixth rule is the entire subject of this piece, and we will come back to it. First, the two controlled languages currently running at the same time.

The first is ASD-STE100 Simplified Technical English, the descendant of Basic English that aerospace adopted and never let go of. Fifty-three writing rules, a dictionary of roughly 900 approved words, most permitted in exactly one sense and one part of speech. Issue 9 came out in January 2025 and elevated it from specification to international standard. It exists because European airlines, most of them not from English-speaking countries, wanted maintenance documentation that could not be misread by a technician at three in the morning with no author to call.

The second has no name, no committee, and no dictionary. It is what a language model does to your prose when you ask it to clean things up. A paper published this week in Nature Human Behaviour is the most careful measurement of it so far. Two days after it appeared, Stanley Druckenmiller gave us the field demonstration.

On Monday the Wall Street Journal published "Let the Bond Market Speak," an op-ed under Stanley Druckenmiller's name arguing that Treasury Secretary Scott Bessent should stop doubling debt buybacks to suppress long-term yields and address the primary deficit instead. Bessent worked under Druckenmiller at Soros. The piece had teeth and the finance world read it.

It also had a certain cadence. Within hours readers were running it through Pangram, the AI detection tool, and posting the results. Commenters flagged the specific construction that gave it away, the one where a sentence denies a thing in order to assert its replacement. Not liquidity management, price management. Not a crisis, an invoice. The economist Claudia Sahm posted that the whole text scored as machine-written.

Druckenmiller confirmed it immediately and without embarrassment. He told NOTUS that of course he used AI, that he moved from an English major to an economics major for a reason, and that he now writes everything this way for the same reason he uses a calculator for arithmetic. He is 73. He named Claude and ChatGPT. He said he rejected many of the suggestions and that the ideas were ones he had been arguing publicly for fifteen years. Paul Gigot, the paper's editorial page editor, backed him, called AI a fact of modern life, and noted that nobody doubts the opinion is genuinely Druckenmiller's. There was no disclosure on the piece.

Set this against the Nature paper and something clicks that neither one shows alone.

The paper's finding is that polishing subtracts the author's markers while preserving meaning, and that readers cannot detect the subtraction. Druckenmiller looks like a refutation, because readers detected it in an afternoon. It is not a refutation. What those readers detected was not the absence of Druckenmiller. Nobody on finance X has a baseline for Druckenmiller's prose rhythm. What they detected was the presence of something else, a set of tics belonging to the model rather than the man.

So the paper measures the subtraction and the op-ed demonstrates the addition, and they are the same event seen from two ends. The signature does not vanish. It gets replaced with a signature that millions of other documents are also wearing. That is what makes it findable, and it is why detection works at all.

Two details make this worse than a gotcha. First, the same thing happened this month to a Harvard economist writing about tariffs in the Financial Times, so this is a pattern and not a stumble. Second, and more uncomfortably for me: the construction that convicted Druckenmiller is on the banned list in my own style guide. I have been running a manual detector on myself for two years without calling it that.

In April 2024 the same lab posted a preprint covering what became studies two and three here. Its conclusion was that LLM involvement slightly reduces the predictive power of linguistic markers, that significant changes were infrequent, and that the predictive power was not fully diminished. Sixteen months later, with new time-series work bolted onto the front and the same core result underneath, the title is The Shrinking Landscape of Linguistic Diversity. The arXiv listing flags the text overlap itself.

Nothing improper happened. The new studies are real, the peer review was serious, and the reviewers signed their names. But the distance between "slightly, and infrequently" and "shrinking landscape" is not distance the data traveled. It is distance the prose traveled, and it traveled it in exactly the direction that gets a paper into a Nature journal. A paper about language quietly conforming to the dominant register of its environment is itself an example of language quietly conforming to the dominant register of its environment. I am not sure the authors would disagree.

I should declare an interest, since the argument implicates me twice over. I have written about AI homogenizing written output more than once here, most concretely in the 23 March 2026 issue on the Berkeley and DeepMind work, where pronoun use dropped 40 to 60 percent under LLM revision and adjectives climbed 57 to 90 percent. I am now telling you a peer-reviewed paper confirms a thing I have been asserting for a while. Notice how satisfying I find that. Then notice that I write this newsletter with an AI, that Druckenmiller named the same model I use, and that if the paper is right you cannot fully tell from the outside how much of this paragraph is me.

Set all three systems side by side and the comparison does something none of them does alone. All three reduce linguistic variance. All three work. Only two of them told anybody, and only one has an escape hatch.

The middle column is the answer to a question the other two cannot resolve. STE has fifty-three rules and no rule six, which is correct, because a maintenance manual has no interiority to protect and a technician improvising at three in the morning is the hazard rather than the goal. Its measured benefit is real and modest: comprehension error rate falling from 18 percent to 14 percent across sixteen workcards and 175 working technicians, concentrated in hard documents and non-native readers. That is what a cage buys you when you build it on purpose and put it around exactly one kind of writing.

The second controlled language has no rules to break, which is a different problem and a worse one. You cannot override a constraint you cannot see. Druckenmiller did not decide to write in the model's register. He accepted help, and the register came with it.

Orwell is the only one of the three who wrote down a constraint and then, in the same breath, wrote down when to violate it. That is not weakness in the rule set. It is the feature that keeps a rule set from becoming Newspeak, and he knew it, having spent the previous four years advocating for a rule set that had no such clause.

Tom Rachman, a novelist who now writes on AI policy, published five rules for AI writing in June. They are ethical rather than mechanical, which is what makes them useful here. Paraphrasing rather than reproducing them: assume your use will be found out; ask whether you would owe a human credit for the same help; recognize that letting the machine write costs you the wisdom you would have earned by writing; expect nobody to read what you obviously outsourced; and if either you or the reader would care whose name is on it, do not generate it.

His first rule is dated 30 June. Druckenmiller was identified by a detection tool eight weeks later. Rachman's underlying claim is that detection got good, and the evidence supports him more than most people assume: a University of Chicago evaluation of roughly 2,000 human and 2,000 machine passages found essentially zero false positive and false negative rates for Pangram on medium and long passages, degrading below about fifty words. Pangram reports targeting no more than one wrongful accusation in ten thousand. That number deserves scrutiny rather than applause, because a false accusation ends a career and the appeal process does not exist yet, but the tool is not the coin flip people assume.

Here is my own set, which is what you asked for. It sits between the mechanism of Simplified Technical English and the license of just writing whatever you want, and it steals its last rule outright.

1. Declare the scope before the first draft, not after. Mark which sentences carry claims and which carry you. Claim sentences can be polished by anything. Voice sentences are not eligible. The whole achievement of STE is a scope statement, and a scope statement takes one line to write.
2. Do the thinking in your own words, badly, first. The model will hand you something better than your bad draft. That is not the point. The bad draft is where the next idea comes from, and skipping it borrows against an account you cannot repay.
3. Apply the credit test to every sentence you did not write. If a colleague handed you this sentence, would you owe them a mention? If the answer is yes, you owe the reader a disclosure. This is Rachman's rule and it is the only one that survives contact with a lawyer.
4. Protect your errors. The sentence that runs long, the word you overuse, the joke that half-lands. These are precisely the markers the Nature paper watched get sanded off. They are not flaws awaiting cleanup. They are the evidence that a person was here.
5. Break any of the above sooner than publish something dead. Orwell's sixth rule, restated. A rule set without an override is not a style. It is a specification, and a specification applied to a personal essay produces exactly what everyone spotted in the Wall Street Journal on Monday.

The paper's own framing points at diagnosis, and it is not a stretch. Linguistic markers are used in research settings to track cognitive decline, to characterize depression, and in screening work on suicide risk. Those methods depend on the same individual variance the paper watched disappear. The contribution is not that these tools will break. It is that they degrade quietly and in a specific direction, which is worse than breaking, because a broken instrument announces itself.

The operational version in health systems has nothing to do with research instruments. Patient-portal messages, intake narratives, and increasingly the clinical note itself now pass through a polishing layer somewhere between the person and the chart. Whatever signal a clinician's ear was picking up from how something was phrased is being normalized before it arrives, and no field in the record marks that it happened. I do not know how large that effect is. Nobody does. It is the study I would want somebody to run.

The useful move is the one aerospace already made, and it is rule one above. Decide where the controlled language applies. Machine alarm text, technician procedures, and medication instructions are the aviation case almost exactly, and they should be as flat and unambiguous as STE can make them. A patient describing what is wrong is the opposite case, and should be preserved as written, with any polished version stored alongside rather than on top.