ML Researchers Forecast Task Parity by 2047, Full Automation by 2116

Artificial intelligence researchers place the median probability of machines outperforming humans at every individual task at 2047.

The finding emerges from a survey of 2,778 authors who recently published in primary machine learning venues. This projected date represents a notable acceleration compared to previous rounds of expert elicitation. Earlier surveys positioned parity decades further out. However, recent advances in generative architectures and large-scale model training have visibly compressed expectations among domain specialists.

Yet the median figure obscures a stark forecasting divide across the research community. The response distribution exhibits immense variance, with individual projections ranging from the near term to several centuries into the future. While aggressive forecasters expect models to master all cognitive tasks before 2035, conservative respondents argue that current techniques will encounter fundamental scaling limits that stall general progress.

The survey also draws a clear distinction between technical parity on isolated tasks and the complete automation of human labour. When asked when all existing occupations could become entirely automated, the median estimate shifted to 2116. This gap of nearly seven decades suggests researchers recognise critical friction between algorithmic competence and practical economic integration. Translating theoretical model capabilities into embodied systems, reliable infrastructure, and commercial workflows remains an enduring operational barrier.

Specific task milestones follow a similarly uneven path. Creative and analytical capabilities, such as writing coherent prose or composing music, are projected to fall to automation within the next decade. In contrast, complex physical operations, including autonomous surgery and dexterous robotics, face substantially longer horizons.

The findings also captured persistent unease regarding downstream deployment. A substantial share of surveyed researchers assigned non-zero probabilities to catastrophic outcomes stemming from advanced autonomous systems.

Can developers establish verifiable safety frameworks before commercial deployment overtakes their compressed forecast horizon?