Generative AI's family tree reaches back to a 1906 poem analysis
Chatbots seemed to appear from nowhere in late 2022, but machines producing text have a surprisingly long pedigree. It starts with a Russian mathematician counting vowels and consonants in Pushkin's Eugene Onegin, passes through a painting program in the 1970s, and only much later arrives at transformers.
In 1906 Andrey Markov described what became known as the Markov chain, testing it on the pattern of vowels and consonants in Pushkin's verse novel. The trick is statistical: learn how likely each symbol is to follow another, then roll the dice to produce fresh sequences. Trained on a body of writing, such a chain can spit out plausible if shallow text, and the idea has been used to model language ever since.
Artists pushed further by the early 1970s. Harold Cohen built AARON, a program that made paintings on its own, and exhibited the results. Meanwhile the word generative meant something else entirely in the 1980s and 1990s, when generative planning systems used symbolic methods to work out sequences of actions, producing military crisis plans, factory process plans and decisions for prototype autonomous spacecraft.
Neural networks took over in the late 2000s, but at first they mostly classified things rather than creating them, because generative models were harder to train. That changed in 2014 with variational autoencoders and generative adversarial networks, which could model images convincingly. A 2017 paper, Attention Is All You Need, argued that the transformer architecture would beat older LSTM networks, and OpenAI's GPT-1 followed in 2018. Public tools arrived in a rush: the voice generator 15.ai in March 2020, DALL-E in 2021, Midjourney and Stable Diffusion in 2022, and ChatGPT that November.
The boom has been uneven. A 2023 survey found 83 percent of Chinese respondents using the technology against 65 percent in the United States, and by mid-2025 analysts at Gartner and The Economist said businesses had slid into the trough of disillusionment as pilot projects stalled. A Federal Reserve survey that October found one in four American workers had used it in the previous month, and 81 percent of those users said it saved them time. Concerns remain over deepfakes, training on copyrighted work without permission, and the water and power consumed by data centres.
Source: Generative AI
