Survey of 2,700 AI Researchers Compresses Expected AGI Timelines

A survey of more than 2,700 artificial intelligence researchers indicates that technical specialists are compressing their timelines for the arrival of human-level machine intelligence.

The inquiry gathered forecasts from peer-reviewed practitioners to assess the trajectory of frontier machine learning. Researchers provided explicit probability distributions over when systems would achieve parity with human workers across varied cognitive and physical tasks.

The resulting dataset highlights a marked acceleration in expected arrivals. In prior iterations of large-scale expert polls, the median expectation for high-level machine intelligence typically hovered in the mid-to-late twenty-first century. The latest findings show that researchers have shifted considerable probability mass forward, with many placing transformative benchmarks within the next two decades.

Respondents attributed this timeline compression to the compounding effects of pre-training scale, multi-modal convergence, and automated feedback loops. Rather than viewing human-level intelligence as an abstract, distant paradigm, technical specialists increasingly model it as the natural continuation of empirical scaling trends.

For the technical safety community, shorter arrival horizons intensify structural bottlenecks. Critical research tracks—including scalable oversight, mechanistic interpretability, and robust goal stability—require substantial empirical iteration to verify. If arrival dates contract while model scale expands, the window to test safety architectures against near-human systems will prove narrower than anticipated.

The survey also underscores significant variance in how researchers evaluate potential failure modes. While consensus has formed around faster progression schedules, respondents expressed divergent views on the likelihood of catastrophic outcomes. Estimates of severe loss-of-control scenarios spanned a wide range, reflecting an ongoing lack of unified risk modelling across major laboratories and academic institutions.

This disparity leaves the field in a precarious operational state. Capability progress continues to outpace formal safety methodologies, driven by private compute investments and commercial deployment incentives. If the median projection among practitioners continues to move forward, will safety verification frameworks be prepared in time?