Anthropic essay: 'Claude-shaped science' needs human judgment

The essay draws on Schwartz's experience building BootLoops, a toolkit he owns and manages, which Anthropic explicitly notes is not its product. Starting from scattering amplitudes in particle physics, Claude moved existing methods into a common framework and extended them to difficult integrals involving elliptic functions. This set the stage for applying physics techniques to ecology, population genetics, economics, and linguistics — fields where the same equations often appear. In ecology, the AI's calculation produced a striking 4.5-fold faster change than neutral theory predicted, but experts cautioned it would not surprise many ecologists, prompting a reframing of the question. Similarly, in population genetics, initial results were not decisive, so the analysis was extended to correlations between variants on the same chromosome, examining 5.7 billion nearby variant pairs from the 1000 Genomes Project and reportedly finding evidence for gene conversion. In economics, an NBER working paper from September 2026 described an AI workflow that examined 4,452 replication packages from five major journals, flagging inconsistencies in 3,460, cutting computation time to under one-tenth in 496 while maintaining accuracy, and creating extensions in 923. Schwartz reported working on 36 manuscripts across 18 fields with 19 co-authors in three months, while noting Claude's tendency to declare victory early and struggle to judge interest. The proposed safeguards — clear pass criteria, human review, independent verification, and expert evaluation — suggest that as AI increases throughput, the design of human oversight becomes more critical, not less.