From ELIZA to GPT: How the uncanny shaped human-AI interaction's hidden history
In 1961, researchers at Bell Labs programmed the IBM 7094 to sing 'Daisy Bell'. The voice was synthesised, mechanical and deeply strange: almost human, but not quite. It was uncanny – that unsettled feeling we get when something is almost, but not quite, human.
That sensation of almost-but-not-quite-human has followed AI development ever since. But when I talk here about the uncanny in relation to AI, I'm not primarily talking about the strange feeling a human can have when a chatbot says something unexpectedly human (though that’s a huge and fascinating subject itself). I'm interested here in something more structural: how the uncanny actually shapes the way large language models work, and why understanding that can make one better at working with them.
ELIZA and the pattern-matching therapist
This story starts in 1966 with ELIZA, built by Joseph Weizenbaum at MIT. ELIZA used simple pattern matching to simulate a conversation with a psychotherapist. It had no semantic understanding, no memory of previous exchanges, and couldn't learn from interaction. It was, essentially, an elaborate search-and-replace algorithm: it scanned your input, matched it against patterns, and filled in response templates. 'I feel anxious' became 'Why do you feel anxious?'
What concerned Weizenbaum (enough to make him one of AI's earliest and most prominent critics) was that users attributed consciousness, empathy and therapeutic insight to this mechanical text processing. People knew they were talking to a programme, but they projected depth onto it anyway. Psychologist Sherry Turkle framed this differently: not as deception, but as something revealing about human psychology. We are, Turkle suggested, primed for relationship with machines. Her concept of the 'second self' suggested that computers function as evocative objects, mirrors in which we project parts of ourselves.
Both perspectives matter. Weizenbaum's caution resonates with recent reports of LLMs reinforcing users' delusional thinking, while Turkle's insight suggests that this isn't simply a technology problem, but a deeply human one.
The uncanny as mechanism, not just feeling
This is where Freud’s framing of the ‘uncanny’ comes in. In his influential 1919 essay 'The Uncanny', he unpacked the German word heimlich (homely, familiar), tracing how its meaning develops in the direction of ambivalence until it finally coincides with its opposite, unheimlich. In this work, the uncanny emerges not simply when the familiar becomes strange, but when the capacity for strangeness was always already latent in the familiar itself.
That structure is doing something specific in LLMs. When we use chain-of-thought prompting – asking a model to 'think step by step' – the model performs human reasoning patterns: it writes out algebraic steps, works through logical stages, shows its workings. This looks familiar. But what's actually happening underneath is the transformer's statistical prediction process, which is deeply strange in relation to human reasoning.
The key insight is that this isn't merely performance; the model isn’t just performing human reasoning patterns for the user. Each step the model generates shifts the probability distribution for the tokens that follow, so writing out reasoning steps constrains the model's predictions toward accurate problem-solving. The familiar human pattern (algebra, logical steps) is actively shaping the strange machine process. The familiar becomes strange; the strange is guided by the familiar. This is what I call ‘uncanny performance’ – structural rather than incidental. The LLM's performance of human reasoning is not human reasoning, but it's precisely that performance that makes the model's underlying process work.
Cognitive maps over tips and tricks
This matters at a practical level. Chain-of-thought prompting and role prompting are often taught as techniques, tips and tricks for the technology of this moment. That's not wrong, exactly, but it has a shelf life, since LLM technology moves fast and best practices from two years ago are already being superseded.
What doesn't become outdated is understanding why these techniques work. American psychologist Edward Tolman developed the concept of cognitive maps in 1948 (internally represented mental models that enable flexible, transferable behaviour across changing contexts), and that's what understanding the uncanny as an underlying mechanism can give us: not a set of prompt tricks, but a cognitive map of human-AI interaction.
Memorising prompt techniques is like rigid function calling, working well until the context changes. Understanding the mechanism beneath them is more like parameter tuning, adaptive and transferable, and applicable to whatever the technology becomes next. If we understand uncanny performance as underlying human-AI interaction, regardless of technical specifics, we can adapt more fluidly to changes in the technology.
Why the history matters
The patterns ELIZA established in 1966 (humans projecting depth onto pattern-matching, machines performing humanness to work more effectively) didn't disappear over the following six decades. They deepened, and with modern LLMs, they became structural: contemporary models don't just trigger the uncanny response in users, they depend on the uncanny structure to function at their best. Understanding that history is more than just context, it's a way into understanding the mechanism itself.
This post is based on my talk 'From ELIZA to GPT: How the uncanny shaped human-AI interaction's hidden history', given at PyConUK 2025 in Manchester. You can watch the full talk, including live demos of chain-of-thought prompting, here.
