AI agents independently conduct real particle physics measurements on archived data

Researchers used AI systems (OpenAI's Codex and Anthropic's Claude) to perform the complete analysis of a real physics experiment — data collection, statistical corrections, and interpretation — with human physicists directing but not doing the work. This is the first demonstration that AI can handle precision scientific measurement end-to-end, not just assist with pieces of it.

For decades, AI systems have been helpers in science: suggesting analyses, writing code, spotting patterns. This work shows AI can own an entire measurement loop — define the problem, execute the method, correct for systematic errors, and produce results that match human physicist standards. If this scales beyond the controlled setting of archived data with known answers, it suggests the bottleneck in experimental physics may shift from data-taking to interpretation and theory integration, and that teams could shrink by replacing human analyst time with AI labor. The catch: this is still a proof-of-concept using 30-year-old data where the right answer is already known. Real signal emerges only if teams start using this for novel measurements where the answer isn't predetermined.