There’s a 5% chance of AI causing humans to go extinct, say scientists

Stephen Taylor / Alamy Stock Photo

A majority of AI researchers have predicted minimal likelihood of superhuman AI causing human extinction, despite recognizing the non-trivial potential danger, in the largest AI conference survey yet. However, they also acknowledged a degree of uncertainty and disagreement about these risks. The survey asked respondents about future AI technology milestones and their potential impacts, both positive and negative. Nearly 58% of the researchers believe there's a 5% chance of extreme harm from AI, like human extinction.

According to one of the authors at the Machine Intelligence Research Institute in California, the overall belief among AI researchers in a non-negligible risk from advanced AI is more significant than the specific percentage. However, they add, we needn't panic: their forecasting ability has been shown to be align with that of the non-expert public in long-term accuracy.

Compared with last year, the participants this year predicted some AI developments to occur sooner. This aligns with newly launched services like the large language model-based ChatGPT in November 2022. The researchers estimated 50% odds of AI surpassing humans across tasks by 2047 and all occupations being fully automated by 2116, 13 years and 48 years earlier, respectively, than last year's estimates.

However, these optimistic predictions might not be realized, as unpredictability is common in AI breakthroughs. Even without superhuman AI risks, many researchers - over 70% - expressed substantial concern over AI-fuelled scenarios such as deepfakes, public opinion manipulation, weapon engineering, population control and economic inequality. They also feared AI worsening democracy and climate change misinformation. The presence of technology capable of destabilizing democracy was particularly worrisome, with an eye to the upcoming 2024 elections.