The Current State of Autonomous Experimentation As of late 2026, the foundational capability for an AI system to autonomously design, run, and revise human experiments already exists. Recent systems have successfully demonstrated closed-loop research pipelines. For example, auto-psych autonomously designed experiments, recruited human participants via the Prolific API, and iteratively revised explicit cognitive models arxiv.org. Similarly, AutoCog executed multiple fully autonomous online studies with real participants, generating explicit mechanisms and code that outperformed standard decision heuristics on held-out data arxiv.org.

Remaining Hurdles for a Qualifying Discovery While the infrastructure is mature, these precedents fall short of the required criteria. Both auto-psych and AutoCog studied tasks outside the five specified domains (randomness judgments and multi-attribute decision-making, respectively) arxiv.org. Crucially, they succeeded by beating baseline seed models (such as Take-The-Best or literature-derived heuristics) rather than the absolute best published human theories. Reviewers of AutoCog noted its domain was relatively "friendly" and its search remained local, largely rediscovering known principles 2 sources. Additionally, automated adversarial collaboration addressing categorization has only been demonstrated on simulated data, not real human subjects arxiv.org.

Expected Timelines and the Path Forward The transition from multi-attribute choice to one of the qualifying domains, such as risky choice or intertemporal choice, is a straightforward next step. Researchers have strong incentives to point existing autonomous architectures at domains with large established benchmarks, such as choices13k. Given the rapid decrease in compute costs and parallel progress in autonomous reasoning capabilities, early preprints claiming breakthroughs in qualifying domains are highly likely by mid-to-late 2027. The median estimate of early 2028 reflects the time needed for the first genuinely successful, rigorous theory to be formulated and publicly released.

Why the Tail is Long The primary bottleneck—and the reason for a right tail extending to 2031—is the stringent requirement to definitively beat state-of-the-art human theories (e.g., BEAST in risky choice) and survive independent lab replication. Early AI discoveries frequently fail to generalize or may overfit to specific datasets; highly rated AI-generated theories do not always produce more accurate empirical predictions papers.ssrn.com. Furthermore, because the qualifying date is the release date of the eventually replicated theory, early claims that fail strict independent validation will not count, resetting the clock. The broad upper percentiles account for the possibility of systemic hurdles in automated scientific reliability and the inherent lag in independent academic replication.