I trained as an aerospace engineer before I ever ran a marketing campaign. So when Andrej Karpathy told people to make their AI write like an aircraft maintenance manual, I smiled.
Matt Southern covered the post on Search Engine Journal. Karpathy, a founding member of OpenAI, recommends asking language models to explain things in ASD-STE100, the Simplified Technical English standard written for aircraft maintenance documentation. He says models know it well and the output is "a lot more readable".
It sounds like a niche tip. It is actually a statement about where the cost of AI work now sits.
Why aircraft manuals read the way they do
ASD-STE100 exists for a brutal reason. A maintenance technician working a night shift, possibly in their second or third language, has to follow a procedure exactly. Ambiguity there is not a style problem. It is a safety problem.
So the standard is strict. It limits the core vocabulary to roughly 900 approved words, most with a single meaning.
It sets maximum sentence lengths, requires active voice, and puts one instruction in each sentence. "Do not" means do not. There is no room for "you may want to consider".
What struck me in engineering was not that the manuals were simple. It was that they were fast. You could read a procedure once and act on it. Nothing had to be decoded.
Generation is cheap, reading is expensive
Karpathy's broader point is the one marketers should keep. As models do more of the work, he says, more of our own work moves "up the abstractions into oversight and understanding". Humans become the reviewers.
That changes the economics. A model can produce a 2,000-word strategy memo in seconds. A senior person still needs fifteen minutes to read it, find the weak claim and decide whether to trust it. Multiply that across a team and reading time, not generation time, becomes the real cost line.
I called this out when I wrote about the AI efficiency gain nobody audited. The productivity numbers in AI case studies almost never include the hours spent checking the output. Default model prose makes that worse. It is long, hedged, padded with transitions and full of words that sound careful but say little.
A controlled language attacks that cost directly. Short sentences expose a weak claim. One meaning per word removes the room for vague phrasing. Active voice tells you who does what, which is exactly what a reviewer needs to know.
How to use it without becoming a robot
Karpathy himself softens it. He sometimes asks the model to go "80% of the way" toward the standard rather than all the way. That is the right instinct for business writing.
Full STE would make a brand campaign sound like a torque specification. But internal work, the summaries, briefs, analyses, meeting notes and reports that people only read in order to make a decision, benefits from going most of the way.
The practical version is a short block in your system prompts or custom instructions.
Mine reads roughly like this: write in controlled English inspired by ASD-STE100, with a maximum of 20 words per sentence, active voice, one idea per sentence, no hedging words, the conclusion first, and numbers instead of adjectives.
Run it for a week on internal summaries and measure one thing: how long it takes the reader to make a decision. That is the metric that matters, not word count.
Karpathy also suggests going beyond text: asking for a diagram, an HTML page or even a narrated explainer video when the topic is complex. The principle is the same. Choose the format that costs the reader the least time to understand.
The same discipline helps machines read you
There is a second use, and it points the other way. If plain, controlled language helps humans read AI output, it also helps AI read your content.
When a model summarises your product page for a buyer, it compresses everything into a few sentences. Clear claims with specific numbers survive that compression. Vague positioning with three qualifiers does not. I made that argument in make your site legible to machines, and STE is a more rigorous version of the same advice.
Take your three most important pages and rewrite the core claims in controlled English: one fact per sentence, with a named product, a specific capability and a measured result. Then check how AI assistants describe you before and after. That is exactly the kind of change GEOflux.ai (geoflux.ai) is built to track, because it maps not only whether your brand is mentioned but what the answer says about you.
Clarity is a safety system
Aviation learned this the hard way. Most of the language rules in that standard exist because someone misunderstood something once and it mattered.
Business errors are usually less dramatic, but the mechanism is identical. A misread brief, an assumption buried in paragraph six, a summary that sounded confident and was wrong. As AI writes more of what we read, the chance of those errors grows unless the writing is built to be checked.
Writing for the tired reader was always good practice. Now every reader is tired, and half of them are machines.