The core tension centers on the strict requirement for an "unchanged methodology" across at least three occasions spaced twelve months apart. We currently have multiple well-resourced publishers—including OpenAI, Anthropic, and third parties like METR—who have published initial R&D automation metrics in 2026. The runway to December 2029 easily accommodates a three-year publication cycle (e.g., late 2026, late 2027, late 2028). However, maintaining a completely static methodology over a 24-month
Set against related questions, this probability was lowered from 41% to 31% because the rapid capability expansions anticipated across our forecasts make a static, unchanged methodology highly improbable over three years.
The 15% probability reflects the strict causal chain required for this outcome. While it is highly likely that a frontier developer will cross a published AI-research-automation threshold by 2031, that makes it slightly less likely that the developer will formally pause capability development and explicitly attribute the halt to that specific threshold, because developers have strong incentives to adapt safety frameworks and attribute necessary pauses to security incidents rather than automation
Factoring in the high likelihood that AI research automation thresholds will be crossed by 2031 slightly lowered this estimate, as developers have strong incentives to adapt safety frameworks and attribute necessary pauses to security incidents rather than automation metrics.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited