State of Play (October 2026) As of late 2026, AI in the social sciences excels at high-volume replication and routine coding, but has not yet delivered substantive, field-defining discovery. The most impressive applications so far are large-scale audits, such as an LLM workflow checking 4,452 economics replication packages nber.org, where humans designed the workflow and AI performed the checking. Meanwhile, efforts to fully automate research have struggled with quality: Project APE produced 1,000 economics papers that remain unreviewed, with an automated audit finding an average of 4.3 issues per paper, 97% of which were confirmed by human reviewers ape.socialcatalystlab.org. Closed-loop systems like AutoCog and auto-psych show promise by running online studies and identifying new cognitive regularities arxiv.org, but these remain narrow proofs of concept rather than top-of-year results.

Pathways to Fast Breakthroughs Despite these current limitations, an early breakthrough by 2027 or 2028 is highly plausible because humans are explicitly permitted to pose problems, execute experiments, and write the final papers. The AI only needs to supply the central idea and the main result. Given the staggering recent pace of AI progress in mathematics—where agents moved from incremental constructions to paradigm-shifting conjectures in about 18 months—a frontier lab could deliberately target social science. The most likely fast routes avoid physical-world bottlenecks entirely: an AI-solved open problem in economic theory or social choice, an AI-discovered natural experiment mined from archival or administrative data, or a scaled-up closed-loop online psychology experiment that settles a major debate.

Bottlenecks and the Importance Threshold However, social science fundamentally differs from mathematics, primarily because it lacks a formal, automated verification environment like Lean. Most paradigm-shifting social science discoveries—such as those challenging theories of European state formation or separating selection from treatment effects in remote work 2 sources—rely on novel real-world data, fieldwork, or deep institutional knowledge. Current autonomous agents exhibit weak scientific taste and often overinterpret noise arxiv.org, a critical liability when navigating messy empirical datasets. Furthermore, the METR benchmark itself cautions against extrapolating performance on clean software tasks to the messy realities of open-ended research metr.org. This points to a median expectation of mid-to-late 2028, reflecting the time needed to move from routine automation to substantively important empirical discoveries.

Attribution and Recognition Frictions The long right tail extending into the early 2030s is driven by severe sociological and institutional hurdles. To qualify, a result must be broadly ranked among the year's top substantive advances in its field, not merely covered as an impressive AI demonstration. That consensus forms slowly and is frequently contested. Even if an AI agent rapidly searches existing datasets to find a brilliant identification strategy, humans must often provide extensive conceptual correction and experimental infrastructure. This blurring of lines makes it highly likely that human authors will simply claim credit for the central idea, undercutting the strict attribution requirement and significantly delaying a clearly recognized, AI-originated top result.