The Hidden History: Why Modern Generative AI Is Older Than You Think

The recent explosion of tools like ChatGPT has fueled a narrative that GenAI is a sudden, recent phenomenon. This idea, however, is a “historical fiction” that confuses the popularization of a technology with the invention of its core components.

In reality, the foundational breakthroughs powering today’s models were developed well before 2018, built upon decades of prior research. Understanding this history is critical for correctly interpreting the legal and policy frameworks that govern AI today.As pointed out by Knowledge Rights 21, “AI text and music generation, automated translation and other generative applications have been around for well over 35 years.”

Standing on the Shoulders of Giants: A Deeper History

The intellectual lineage of GenAI is long and rich.

• 1960s: Basic forms of GenAI were already emerging. ELIZA, a simple chatbot built in the 1960s, could simulate conversation.
• 1980s: Researchers were experimenting with algorithmic music composition and automated text generation. Policymakers were already grappling with the implications; the UK introduced a legal provision for computer-generated works into its copyright law in 1988, reflecting an awareness of these emerging technologies.
• 2010s: This decade saw a rapid evolution. Variational Autoencoders (VAEs) emerged in 2013, and by 2015, “robot journalists” were already auto-writing basic news pieces from data..

The Architectural Pillars of the Current AI Boom

Two key technologies, both established before 2018, served as the direct ancestors of today’s GenAI wave.

1. Generative Adversarial Networks (GANs) – 2014 

The GAN concept was introduced in a groundbreaking 2014 paper by Ian Goodfellow and his colleagues. It works by pitting two neural networks against each other: a “generator” that creates new images and a “discriminator” that tries to distinguish the fake images from real ones. This adversarial process was a massive breakthrough, enabling the generation of high-quality, realistic images nearly indistinguishable from human-created ones. The code was made public, spurring a wave of research long before 2018.

2. The Transformer Architecture – 2017 

The second, more critical, pillar is the Transformer architecture, as detailed in the 2017 Google paper “Attention Is All You Need.” Its key innovation was the “attention mechanism,” which allowed the model to weigh the importance of different words in a sentence all at once, rather than sequentially. This enabled parallel processing, unlocking the ability to train models on vastly larger datasets than ever before. This development led directly to modern Large Language Models (LLMs), including OpenAI’s first Generative Pre-trained Transformer (GPT-1), which was developed in 2018, followed by GPT-2 in 2019.

Why This History Is Critical for Law and Policy

The assertion that “generative AI did not exist in 2018” is a fallacy that confuses a consumer product with the underlying technology’s existence. It’s like arguing the internal combustion engine didn’t exist until the Ford Model T was mass-produced.

The core scientific and engineering work that enables systems like ChatGPT and Midjourney was completed and published in 2014 and 2017. By 2018, GenAI was already making headlines. A famous example is the “Portrait of Edmond de Belamy,” an artwork generated by a GAN, which sold at Christie’s auction house for $432,500. The argument is therefore not about the technology’s existence but about its mainstream visibility and commercial packaging. 

This distinction is not merely academic. The false narrative of AI as a sudden disruption is often used to argue that existing laws could not have anticipated it. The historical facts prove otherwise as we will see in part 3 of this series. Legal frameworks like the EU’s 2019 Copyright Directive were drafted and debated during the very period when these foundational technologies were known and maturing. This context is essential for correctly interpreting the law’s intent and application to AI.

Written by Caroline De Cock, LL.M. , Head of Research.