2022: the year the public arrived

This was the year AI stopped being something most people only read about. Two threads converged. The first was images: OpenAI’s DALL·E 2, the Midjourney beta and, most consequentially, the open-source release of Stable Diffusion, which put text-to-image generation on ordinary computers rather than behind a company’s paywall, and within weeks it was everywhere. The second was conversation. Quiet groundwork earlier in the year — teaching models to follow instructions with human feedback, and the discovery that simply asking a model to reason “step by step” improved its answers — fed into the launch of ChatGPT at the end of November. It reached an estimated hundred million users within two months, one of the fastest adoptions of any consumer product, and it did more than any paper to make the capabilities of large language models legible to a general audience. Underneath, the research also pointed forward: early work on chained reasoning and on models that could take actions foreshadowed the agents to come.

The same releases carried the year’s frictions in with them. Stable Diffusion’s openness, which its supporters prized, also meant it had been trained on billions of images scraped without consent, and it triggered the first sustained backlash from artists and the copyright disputes that followed. In October the United States restricted exports of advanced chips to China, opening a geopolitical front that has shaped the industry ever since. And ChatGPT’s very fluency turned out to be double-edged: it stated wrong things with the same confidence as right ones, a habit soon labelled “hallucination”, and schools, newsrooms and employers began worrying almost immediately about misinformation, cheating and the difference between a system that sounds authoritative and one that is.