Forecasting AI timelines

Office workers, office, admin, administration, clerk, accountant, customer service, call centre, assistant, Programmers, programmer, developer, software engineer, coder, coding, IT

Writers and journalists, writer, journalist, author, editor, translator, media, news, Artists and designers, artist, designer, illustrator, musician, photographer, actor, creative, Scientists and researchers, scientist, researcher, research, academic, laboratory, lab

Estimating when AI will match or overtake humans at most work, and how fast it will get there.

How much time is there to make AI safe? That depends on when AI becomes capable enough to be dangerous, so forecasting it has become a field of its own. The AI Safety Atlas focuses on two methods: following trends in the computing power, algorithms and data used to train AI, which have grown fast and steadily, and comparing that computing power with estimates of what the human brain does. Other approaches are surveys of experts and forecasting platforms, such as Metaculus, where people make predictions and are scored on how accurate they turn out.

The answers have been moving closer. In a 2023 survey of 2,778 AI researchers, the combined forecast put a 50% chance on machines outperforming humans at every task by 2047, 13 years earlier than the same survey a year before. Some forecasters expect it much sooner: the authors of AI 2027 described 2027 as their single most likely year when they published in 2025, though their median estimates were later.

Every forecast is uncertain; the Atlas stresses that “all forecasts come with wide error margins that compound over time”. But uncertainty cuts both ways. If powerful AI could plausibly arrive within a decade, preparing for it can’t wait until it’s certain.