AI is here because it was built out of human language, the system people already run on, and that is both why it works and why it carries our own flaws with it.
They tried writing the rules by hand from 1954, and the money stopped in 1966. What worked was a count. In 1913 Andrey Markov guessed the next letter of a Pushkin poem with no grammar at all. That move, run over everything people have written, is what sits underneath every chatbot.
A word reaches your body before you agree to it. A banknote is money, a marriage a marriage, only while people accept the words that say so.
Your motor cortex is the strip of brain that moves your body. In a scanner study, reading kick, a leg word, raised the response in its leg region. Reading pick, an arm word, raised it in the arm region. Language is not a tool people pick up. It is what they run on: Yuval Noah Harari calls it the operating system of civilisation, because money, law and religion are built from words. Almost all of it is written down.
The study’s three word classes
Hover or tap a word.
Passive reading, fMRI, 2004. Each word class raised the response in the motor region for the body part it names. This is one study, and how much of the motor strip a word really engages is still argued; the section rests on it alone.
as of 2026-08-26 · Hauk, Johnsrude & Pulvermüller, Neuron 41:301–307, 2004 · one study, and how much of the motor strip a word engages is still argued
The obvious way to give a machine the system is to hand over the rules. Nobody had settled what they were.
For most of the last century linguists have held that every language shares rules children are born knowing. What those rules are is still argued. The live case is Pirahã, an Amazonian language of a few hundred speakers. It appears to do without the sentences inside sentences every language was supposed to have. So from 1954 the engineers wrote rules of their own: Russian into English, for a government that wanted to read Soviet documents, one hand-written rule at a time. The rules multiplied faster than anyone could write them. In 1966 a US committee, ALPAC, found machine translation not worth its cost, and the funding was sharply cut.
as of 2026-08-26 · Jones 1786 to Pullum 2024, six primaries, in the table below · the sources on Pirahã are split
| Year | Series | Event |
|---|---|---|
| 1786 | The argument | Jones: a judge in Calcutta saw that Sanskrit, Greek and Latin were relatives |
| 1954 | The attempt | Georgetown–IBM demonstration: hand-written rules, Russian into English |
| 1966 | The attempt | ALPAC: machine translation had not met expectations and was not worth its cost; the funding was sharply cut |
| 2005 | The argument | Everett on Pirahã |
| 2012 | The argument | Piantadosi et al. |
| 2024 | The argument | Pullum |
| 240 years | The argument | 1786 to 2026, open (2026 − 1786) |
| 12 years | The attempt | 1954 to 1966, closed (1966 − 1954) |
Nobody ever wrote down what language is. So nobody could hand it over.
what worked instead · a count, 1913
Instead the engineers handed over the record: the system, written down.
In 1913 Andrey Markov wanted to win an argument. A rival held that probability works only on events that do not touch, like coin flips. Markov took a text, where each letter leans on the last, and counted the letters of Pushkin's Eugene Onegin by hand, vowel or consonant. Each letter changed the odds on the next: after a vowel, usually a consonant. The averages still settled where probability said they would. Markov won, and found a way to guess at language by counting it.
A century later the machines were big enough to count the whole record. The same move ran over all of it and left the machine holding the odds: after these words, this word, this often. Show it a text and it looks at every word so far. It picks the one the record makes likeliest, writes it down, and looks again. The move itself looks nothing up and consults no meaning. A product built on it can be handed search and tools, but this is the move underneath. Your question is the text, and what follows a question in the record is an answer.
A. A. Markov, by hand, 1913
Pushkin, Eugene Onegin, opening lines, spaces and punctuation removed, as Markov did. Each stone is one letter, in the order it was written.
the poem’s first line, stone by stone25 of 20,000 letters
Each letter changes the odds on the next.
20,000 letters of one poem, classified by hand, vowel or consonant. 23 January 1913 · Imperial Academy of Sciences, St Petersburg
Markov did it with letters, by hand, in 1913. The machine does it with words, over the written record, today. It is one move.
as of 2026-08-26 · Markov, 1913, via B. Hayes, American Scientist, 2013 · the count read from the primary, the 8,638 / 11,362 split from a secondary
| What | Value | Unit | Basis |
|---|---|---|---|
| The source | 1913 | A. A. Markov, address to the Imperial Academy of Sciences, St Petersburg, 23 January 1913, on the first 20,000 letters of Pushkin's Eugene Onegin, against Nekrasov; via B. Hayes, “First Links in the Markov Chain”, American Scientist, 2013. | the primary, via a scholarly secondary |
| Letters counted | 20,000 | letters | read from the primary |
| Vowels | 8,638 | letters | from a secondary |
| Consonants | 11,362 | letters | from a secondary |
| Vowel share | 43.2% | of letters | derived from the two rows above |
| Date the count was presented | 23 January 1913 | date | read from the primary |
| Letters shown on the rail | 25 | letters | counted on the rail above |
| Vowels / consonants on the rail | 9 / 16 | letters | counted on the rail above |
Before it, machines had to be met in their own language.
In 2023 the AI researcher Andrej Karpathy wrote that the new programming language is English. In 2025 he added a third rung, Software 3.0, to the two he had named in 2017. Software 1.0 was instructions only programmers could write. Software 2.0 was the machine that read the record and came out holding the odds. Software 3.0 is telling that machine, in a sentence, what you want. Anyone who can write can operate it.
Three ways to give the same instruction. Hold a rung to read what it is.
Illustration; the rungs are Karpathy’s, the example is ours. He named the first two rungs in 2017 and added the third in 2025.
as of 2026-08-26 · Karpathy, “Software 2.0”, 2017; post, January 2023; “Software 3.0”, Y Combinator AI Startup School, 17 June 2025
In 5.8 million decisions by professional Go players between 1950 and 2021, decision quality rose after a program that plays above human level arrived. Moves nobody had played before became more frequent.
What makes it useful is what makes it costly. It answers with the commonest answer.
The people who took their answers from one machine reached further, and the room around them went quiet.
The likeliest next word is the commonest in the record, so its first answer is the common answer. Ericsson gave its staff a model. In 18 interviews in 2024, workers there said the model gave them expertise outside their own field, and that they asked each other less. Everyone reaches wider, everyone reaches the same way, and what one person knew stops reaching the next. The cost arrives dressed as an improvement.
The Generated Expertise · Uppsala University, 2024 · with Ericsson
Workers described getting usable answers in fields that were not their own, without asking anyone who worked in them.
the study’s own name for it: hybrid expertise
The same workers described asking their colleagues less, so what one person knew reached the next one less often.
the study’s own name for it: the silo effect
as of 2026-08-26 · Knutsson & Viklander, “The Generated Expertise”, Uppsala University, 2024, 18 interviews at Ericsson. The study names the two themes but does not publish how many of the eighteen described each, so the columns are not sized
The second flaw is what it cannot do about the first. It has never touched anything. A word is grounded when it is tied to the world. Hot is grounded for you because you have been burned. The machine's words are tied only to other words, as Bender, Gebru, McMillan-Major and Shmitchell argued in 2021. Its grip on the world is second-hand, through us. Where the record is wrong, the machine is wrong with it, confidently, with no way to notice.
It gives the common view with the same confidence as the correct one. We never say it understands.
We are two people, and we ship products because the machine takes instructions in the language we speak. MagicLens, Momentos and Studio Pro are built that way. It answers with the common answer, so everything we make starts from the user's own words and keeps them. It cannot tell us when it is wrong, so a person signs off before anything reaches anyone.