Anthropic Models Now Filter 80 Percent of New Biology Datasets Before Human Review

The Atlantic piece frames the advance as a neutral productivity gain. In practice the models are trained on proprietary datasets controlled by the same labs that receive Anthropic compute grants, creating a closed loop where only questions compatible with existing model architectures receive funding. Labs without those grants report their grant applications rejected for lacking 'AI readiness' language.

Traditional biology departments face an incentive shift: principal investigators now allocate 30-40 percent of new postdoc slots to prompt engineers rather than bench scientists. This reallocates NIH and NSF budgets toward compute credits that flow back to the model providers, documented in 2025 award summaries from the National Institute of General Medical Sciences.

The structural outcome is consolidation. Smaller university cores lose access to raw data streams as journals increasingly require AI-preprocessed figures. Within two years the median biology paper may cite model version numbers the way earlier papers cited restriction enzymes.

Next, expect journals to formalize AI co-authorship rules by late 2027, locking in the division between data owners and data interpreters.