probability: 15%

rationale:

Resolution Requirements and Status Quo. A positive resolution demands a strict conjunction of events by the end of 2031: a frontier developer must publicly declare its own AI-research-automation threshold reached, and it must publicly report a halt or pause in capability development because of that threshold. As of late 2026, no developer is near these limits. Anthropic explicitly notes that it has not seen a sustained 2x acceleration in its pace of progress anthropic.com, and OpenAI determined that its Astra-class models fall short of even the "High" tier for AI Self-Improvement openai.com. While internal metrics show rising agent use, evidence suggests that high-level planning and autonomous research remain a minimal fraction of agent outputs openai.com. Note that the July 2026 OpenAI reinforcement-learning pause was a response to a security and misalignment incident, explicitly disconnected from AI-R&D thresholds openai.com, and does not count toward resolution.

The Frameworks and "Drift." Even if a developer declares an AI-research-automation threshold crossed, a development pause is far from guaranteed. Among the leading labs, only OpenAI's framework contains an explicit trigger: "Until we have specified safeguards and security controls that would meet a Critical standard, halt further development" cdn.openai.com. However, this gives OpenAI a strong incentive to pre-specify those safeguards to avoid a halt altogether. In contrast, Anthropic and Google DeepMind frameworks mandate safeguards, expanded security protocols, safety cases, and reporting, rather than automatic capability-development stops 44 sources. Furthermore, there is a documented trend of framework drift, where early pause pledges are replaced by competitor-contingent, discretionary language, or active rewrites as thresholds approach.

Incentives Against Unilateral Pauses. The moment a lab crosses its AI-R&D automation threshold is precisely when its developmental lead compounds fastest. Pausing capability development at this stage would be enormously expensive. Competitive and commercial pressures heavily cut against unilateral stops, and current governance structures provide no backstop—no jurisdiction legally restricts AI as a research instrument. If a developer reaches this threshold, the most likely outcome is that it applies heightened security controls, claims acceptable residual risk, and continues training, rather than publicly freezing its capability scaling.

Decomposition and Final Estimate. The probability that a frontier developer publicly declares an AI-research-automation threshold reached by 2031 is substantial—roughly 40-50%—given the steep capability trajectory and the five-year horizon. However, the conditional probability of a publicly reported development pause attributed specifically to that threshold is only around 30%. While OpenAI has demonstrated a willingness to publicly disclose short, targeted pauses in response to other internal triggers (such as the July 2026 cyber-security pause openai.com), labs are highly incentivized to prepare for R&D thresholds in advance by operationalizing safeguards or revising their safety frameworks to avoid a hard stop. Multiplying these factors yields an overall probability of 15%.