Which of these will happen first?

rationale:

As of mid-October 2026, none of these three triggers has occurred. While AI's role in AI research is expanding rapidly—Anthropic reports that over 80% of its merged code is authored by Claude and that the model operates at or above a "collaborating" level on more than 90% of AI R&D work https://www.anthropic.com/institute/measuring-pace-of-ai-development—no measured subset is fully autonomous https://www.anthropic.com/institute/measuring-pace-of-ai-development. OpenAI has announced reaching an "automated research intern," but stresses that humans still set priorities, steer tasks, and make deployment decisions https://openai.com/index/research-acceleration-view-inside-openai/. The 5.2-year horizon to 2032 gives ample time for capabilities to advance, but the strictness of the resolution criteria and the changing political and regulatory climate strongly shape the likelihood of any official claims.

An official claim of full AI research coding automation is the most likely first event (52%). Coding is the most mature segment of the research pipeline, and the broader picture indicates a steady progression where labs will formally announce the automation of specific sub-categories on the path toward replacing all coders . Expectations that labs will announce human obsolescence for specific categories by 2030 make a qualifying coding claim before 2032 more likely than not. Making a claim restricted to coding alone does not trip any of the leading labs' critical safety-framework thresholds. Therefore, the institutional cost of making this claim is low. Furthermore, a statement that a lab's engineers "no longer write code" holds significant commercial and recruiting appeal, particularly for competitive developers like xAI or OpenAI. What keeps this only slightly above 50% is the strictness of the "requires no human coders" criterion. Even as AI authors nearly all code, human review, testing, and steering remain persistent bottlenecks https://www.anthropic.com/institute/recursive-self-improvement, and labs face liability and regulatory pressures to emphasize keeping "humans in the loop" Top engineers at Anthropic, OpenAI say AI now writes 100 ....

A claim of full AI research automation is highly unlikely to resolve first (4%). By definition, full research automation includes coding, meaning this event would either have to coincide with or inexplicably beat a coding-specific announcement. More importantly, making an official claim that human researchers are no longer needed at any stage would be an admission of crossing the highest-risk thresholds in the labs' own safety frameworks. Under OpenAI's framework, for example, crossing into fully automated AI R&D commits the company to halting development The AI policy window is open. We need to act.. Consequently, labs face severe operational and reputational disincentives against making an absolute, official statement of full research automation, even if capabilities approach that level.

A binding government restriction on the use of AI to conduct AI research remains a distant prospect (9%), though political pressure is building. Currently, no such restriction exists; major governance frameworks—including the EU AI Act, California's SB 53, and recent US executive orders—regulate models as products rather than as internal research instruments Promoting Advanced Artificial Intelligence Innovation and Security – The White House. The recent White House Accord is strictly voluntary White House AI Responsibility Accord & Super Intelligence ..., and congressional bills addressing internal model usage remain at the introduced stage https://www.congress.gov/bill/119th-congress/house-bill/9925/text. While growing calls from lab leaders to establish international "speed limits" on recursive self-improvement https://darioamodei.com/post/we-must-pace-the-frontier could eventually yield a treaty or domestic restriction, such rules face high political hurdles and would likely arrive too late in the window to beat a coding automation claim.

There is a substantial probability that none of these events occurs before 2032 (35%). This captures the very real possibility that capabilities plateau just short of complete human substitution, particularly in complex, long-horizon research judgment and system architecture. Furthermore, it accounts for the stringent wording required by the criteria. Labs may increasingly rely on cautious phrasing—such as "essentially all code" or "99% of work"—to describe their systems without ever officially declaring that human coders or researchers are completely unnecessary. Combined with the likelihood that governance remains largely voluntary or focused on public deployment rather than internal R&D, maintaining the status quo is a highly competitive outcome over the next five years.