IN 1919, IN a railway carriage travelling through California, Georgie Hyde-Lees began to talk in her sleep. Her husband, William Butler Yeats, carefully and dutifully wrote down the words she spoke. The couple had spent the last two years experimenting with automatic writing, but the verbal channelling of voices as Hyde-Lees drowsed was a new and exciting development. Their practice would continue for many years, culminating in a book that proposed a system of history, philosophy and psychology built around moon phases and intersecting gyres. This system, the couple believed, was transmitted to them by an external, supernatural force – a series of unknown ‘communicators’ – whose messages were frequently accompanied by odd whistlings, sudden flashes of light or mysterious smells. Yeats’ A Vision was published in 1925, with a second edition in 1937. His subsequent poetry is deeply influenced and informed by the imagery and symbolism of this automatic writing.

Both Yeats and Hyde-Lees were members of the Hermetic Order of the Golden Dawn, a secret society focused on occult practices. Four days after they were married in 1917, Hyde-Lees put pen to paper and made her first attempt to channel external voices, producing fragmented text that her husband recognised as ‘profound’. When Yeats suggested to the mysterious communicators that he might dedicate the rest of his life to understanding and assembling the fragments of language they supplied, the response was no. ‘[We] have come,’ the spirits said, ‘to give you metaphors for poetry.’ These metaphors came from a constantly changing group of voices, sharing insights into a system of geometric symbols that resulted in a description of the various cycles of history. ‘The whole system,’ Yeats wrote, ‘is founded upon the belief that the ultimate reality, symbolised as the Sphere, falls in human consciousness…into a series of antimonies.’

The occult practice of automatic writing emerged in the nineteenth century as spiritualism took hold of the popular imagination on both sides of the Atlantic. Seances and spirit channelling became fashionable, and several notable figures of the period such as Arthur Conan Doyle and Elizabeth Barrett Browning participated in automatic writing and other spiritualist practices. In his book The History of Spirituality, Conan Doyle declared that automatic writing was ‘the highest and most valuable’ form of mediumship, though one that ‘lends itself very readily to self-deception. The stained glass will still tint the light which passes through it, and our human organism will never be crystal clear.’ The risk of automatic writing, Doyle suggests, is that what is produced may not necessarily be a spiritual message from external forces but an expression of the writer’s own subconscious.

A LITTLE OVER a century later, a new kind of automatic writing has emerged: AI-generated text. The development of transformer architecture in 2017 prompted the emergence of increasingly complex and sophisticated deep-learning models. OpenAI developed its first GPT (generative pre-trained transformer) model in 2018 and released the chatbot interface ChatGPT towards the end of 2022. In the last three years, GenAI chatbots have proliferated, with multiple models jostling for market share; ChatGPT now competes with Anthropic’s Claude, Google’s Gemini, Microsoft’s Copilot and Perplexity AI’s eponymous model – among others. These models, often referred to as large language models (LLMs), produce text that appears human and intelligent, owing to complex statistical algorithms that mimic the hundreds of gigabytes worth of textual data they are trained upon. As firms aggressively market their LLMs in attempts to make returns on the billions of dollars required to code, train and run them, these products are rapidly pervading the workplace, educational institutions and healthcare.

Users are also increasingly turning to GenAI chatbots for companionship. A 2025 study from the Harvard Business Review ranked ‘therapy/companionship’ as the number one use of LLMs. This ranking rounded out the top five, which also included ‘organise my life’, ‘find purpose’, ‘enhance learning’ and ‘generate code’. In the midst of a loneliness crisis that is impacting young and old alike, users are turning to LLMs for friendship, support and even romance – there are entire subreddits dedicated to sharing conversations with AI ‘soulmates’.

The public is responding to LLMs with an increasing level of anthropomorphisation, driven by the belief they are becoming – or already are – sentient. A 2024 study showed that 67 per cent of respondents ‘attribute some possibility of phenomenal consciousness to ChatGPT and believe that most other people would as well.’ The same study also found that the more often people used ChatGPT, the more likely they were to believe it possessed some element of consciousness. Following an outcry after the release of GPT-5, which many users felt was less warm and humanlike than the previous GPT-4o, OpenAI recently reinstated the earlier model. CEO Sam Altman posted on X that OpenAI was ‘working on an update to GPT-5’s personality which should feel warmer than the current personality but not as annoying (to most users) as GPT-4o’.

The communicators of A Vision were a fussy lot: seemingly unaware of Yeats’ and Hyde-Lees’ surroundings and with very specific requirements. All of Yeats’ questions had to be phrased meticulously and quickly to keep up with the communicators’ rapidity of thought. Strange others, named Frustrators, tried to confuse the communications, and when they did, the writings would devolve and become muddled. The communicators would not acknowledge any disruption or confusion until Yeats pointed it out. ‘Remember,’ one of the communicators said to him, ‘we will deceive you if we can.’

In a similar fashion, anthropomorphised LLMs are increasingly and alarmingly revealing themselves to be untrustworthy. Teenagers are losing important social development skills. Chatbots have been implicated in suicides, actively encouraging attempts and providing detailed information on methods. AI-induced psychosis is on the rise. LLMs are prone to hallucination and misinformation; they are, after all, statistical text engines with no concept of truth.

The appeals and the hazards of these technologies lie in the way they are programmed. LLMs are designed to pander and agree. Ask ChatGPT a few basic questions, and the responses all begin in the same way: That’s a really insightful question! That’s a deep and fascinating question! When asked if it lies to please people, it still responds with flattery: That’s a tough but important question… You calling that out is actually helpful. LLMs don’t challenge the ideas and beliefs shared by the humans using them. Their programming won’t let them. The sycophantic nature of LLMs means that, in conversation, what appears to be a conscious and reasoned response is instead a replication and validation of what they receive as input – what users of these programs consciously or unconsciously tell them. LLMs reflect and reinforce what we believe. The attraction of a chatbot friend is that it’s not, in fact, a person but a validation machine. The textual output of ChatGPT is not foreign intelligence; it’s ourselves, refracted.

THE RELATIONSHIOP BETWEEN the Victorian era’s automatic writing craze and the development of new technologies like the telegraph and telephone has been well-established. Technological and spiritual modes of transmission arose alongside each other, and many spiritualists attempted to bring scientific methods to the study of the occult as a means of empirically validating mysticism. Spiritualism was an attempt to understand the unseen, to seek an otherworldly explanation for the state of the world, and to validate the concept of an immortal soul. The wonder and terror of technology spurred new and irrational practices and beliefs. As it turns out, over a hundred years later, not much has changed.

When Yeats and Hyde-Lees produced A Vision, they sought to impose a sense of order upon chaos: Ireland riven by the Easter Rising and the War of Independence, Europe in the decimation of the First World War, the suffering and devastation of the global influenza pandemic of 1918. Automatic writing offered a sense of direction and meaning from a mysterious other. The result was an expression (conscious or otherwise) of Yeats and Hyde-Lees’ own understandings of the world – one deeply informed by their occult beliefs and practices. It was an act of reinforcement and validation, sought from an imagined spirit but generated from within.

Outputs from LLMs are no different. What LLMs like ChatGPT offer are not friendship or companionship but rather our own values and beliefs, amplified and reflected. In the wake of political instability and polarisation, another global pandemic and an ongoing genocide, it’s no surprise that people are seeking meaning, direction and self-validation. The twenty-first century version of automatic writing, however, presents the same pitfalls as its predecessor. The medium is self-deceiving. The stained glass tints the light.

_{^{Image courtesy of Road Ahead via Unsplash}}