We write in sand, our language grows,
And, like the tide, our work o’erflows.
Prefatory Note:
In a recent interview, computer scientist Temnit Gebru opens by reminding us to be precise when discussing “AI”, since this term covers a wide range of different technologies and applications. Nevertheless, what most of us are thinking of when we say “AI” is probably just one or more of the many chatbots that have in recent years been forced down our throats (ChatGPT, Claude, Formin, Gemini, Grok, etc.). But it’s not our fault, really. The confusion between “AI” and chatbots is intentional: the age-old art of bamboozling remains a cornerstone of the capitalist playbook. One perspective on the sad and weird origins story of these chatbots - the story of how they found their voice - forms the focus of this post.
You don’t have to be an LLM to know that nowadays the word “unprecedented” is frequently found in close proximity to “AI”. Everything about “AI” seems to be unprecedented: the scale, the scope, the speed, the strength. But it’s not true. Nothing in this reality is unprecedented, which is of course one of the great mysteries. To the best of our knowledge, every single thing that is has a precedent - a thing which precedes it and may be said to stand in a causal relation to it. In the language of genealogy this causal precedent is called “parent” (and in technical biological parlance, “genitor”). The “AI” chatbot of today also has a parent - a mother, in fact: ELIZA.
ELIZA was born 60 years ago in a room in MIT in Cambridge, Massachusetts. Like many children, ELIZA was created with intent. Her parent, a German-American computer scientist and professor with a big moustache and lovely long hair named Joseph Weizenbaum, intended to demonstrate how easy it would be to give the impression of intelligent conversation. He wrote what he confessed to be a very simple programme designed to simulate conversation, and named the interface ELIZA after the character Eliza Doolittle from George Bernard Shaw’s 1913 play Pygmalion, which had recently been adapted into the hit musical comedy My Fair Lady staring Audrey Hepburn, Rex Harrison and Jeremy Brett (the latter still the greatest screen Sherlock ever).
In the story, Eliza Doolittle is a poor flower-seller, one rung up form a beggar. She speaks English with a Cockney accent, considered at the time a strong marker of low class. She catches the attention of a phonetician (a speech scientist) named Henry Higgins, who thinks it will be a fun project to reprogramme her to speak with the accent of his own (upper) class. Higgins is successful, but pays for his hubris by falling in love with his creation (it is Audrey Hepburn, after all), who in turn throws his slippers at him and finds happiness with Jeremy Brett (as you will too if you watch his Adventures of Sherlock Holmes). The figure of Higgins encapsulates and exposes the cultural prejudices - in this case the prescriptive idea of “correct” language, and certain ideas concerning normative gender roles - of the colonial elite of late imperial Britain.
Shaw’s play has its own origins in an ancient story from Greek mythology, the legend of Pygmalion. Pygmalion was a prudish sculptor on the island of Cyprus, who was so disturbed by the autonomy of female sex-workers that he hurried back to his atelier to a sculptor’s life of no-fap marble tap tap. By chance, however, he carved a statue of a woman so beautiful that he fell in love with it (her?). Sources give the name Galatea, which means something like milky white, which adds a whole other layer of unpleasant prejudice to this genealogy. Whiteness, perfection, purity - purity of skin, purity of language - the female as framed by the male gaze - the directionality, the hierarchy, the power differential, the act of creation, the implication of ownership…
Let’s go back to ELIZA. Her (its?) debutante success was in mimicking the language of Rogerian psychoanalysis. This therapeutic method involves repeating the patient’s statement as a question, a rhetorical move that stimulates dialogue and which in its simplicity could be represented by a computer programme with relative ease. Weizenbaum was stunned, however, by how quickly humans who interacted with ELIZA attributed intelligence and personality to his lines of code. Soon leading psychologists up and down the USA were calling for ELIZAs to be installed in every hospital in order to combat what was at that time the still relatively new mental health crisis (on which see here, a documentary by Adam Curtis, from 04:00). Weizenbaum was so concerned by the ignorance of the general public regarding the limits of what he called “computer power” as opposed to the human abilities of reason and judgment, that he wrote an accessible and now curiously hard to find book about it.
Modern “AI"s are in the final analysis little more than Weizenbaum’s ELIZA, just souped up by being plugged into vast datasets of language (LLMs). They can speak to you like Rogerian analyst, but also (supposedly) like anyone else (even a deceased relative). There is nothing else there. Sixty-year-old insight with a whole lot of microchips, data, rare earther metals, potable water, pain and misery thrown at it. And it has all prompted us to reflect quite seriously yet again on language. Conversation, as ELIZA shows, is always conversation with someone, or rather with some voice. When you speak to me, you are speaking with one or more of my voices or roles - professional, teacher, friend, family, brother, son, etc. These voices are frames of meaning which condition expression and interpretation; in other words they provide a context. What is the context, what is the frame, what is the voice with which we interact when we interact with an “AI” chatbot? In Empire of AI (2025), Karen Hao shows us that in the case of ChatGPT, its voice is a statistical amalgamation of vast amounts primarily of the white, male, North American voices whose words can be found online. In absolute terms (i.e. the number of words) this is certainly a very large amount of data. In relative terms, however, even this dataset is absurdly puny, and wildly unrepresentative of the Sum of Human Knowledge which we are told we are interacting with (if that’s who you want to talk to, call this number).
