AI reaches the claims desk first
-
Nº CVII
- Date
- 25 Sep 2026
- Issue
- 107
- Stories
- Seven
- Editor
- ARC
Medical AI's largest live deployment isn't at the bench — it's sitting at the claims desk.
AI systems now help decide who receives medical care
AI now gates coverage decisions at major US health insurers, according to a Vox report circulating on Hacker News that points at UnitedHealth, Cigna and Humana alongside Medicare Advantage plans. The systems involved aren't diagnosing anyone; they score claims and prior-authorization requests, and the report argues that decision volume has outrun meaningful human review. That makes utilization review, not the clinic or the bench, the largest real-world deployment of AI in American medicine. Whatever rules get written in response will set the default expectations for provenance, appeal, and human sign-off that every downstream clinical AI system inherits.
PRAXIS-VirtualCell runs agent experiments on virtual cells
PRAXIS-VirtualCell wraps agent-run experiments on computational cell models in a programmable specification layer, per a new arXiv preprint from Zhenyu Ma and colleagues. Virtual-cell work has drawn serious money on the promise of running perturbations in silico before anything touches a plate. The harder problem is stating precisely what an agent (software that carries out multi-step work on its own) was asked to do, then checking that it did it. Treating that as a design requirement rather than a write-up afterward moves virtual-cell experiments toward results other groups can actually rerun.
OpenAI ships a test for mental-health conversations
OpenAI shipped MentalHealthBench an expert-informed test set scoring how helpful and safe model replies are across realistic mental health conversations. General chatbots already field this traffic at volume with no clinician anywhere in the loop, and the safety claims around it have had no shared scoreboard. A published benchmark turns "our model handles crisis conversations responsibly" from a marketing line into a number competitors can be measured against — and pulls psychiatric safety into the same evaluation culture the rest of clinical AI already lives in.
TopU-LBVS benchmarks ligand screening across multiple targets
TopU-LBVS scores virtual screening across multiple targets under conditions its authors describe as realistic, tightening a benchmark family long criticized for flattering ligand-based models. Hit-rate claims in early discovery now have a harder reference point to clear.
miRstring decodes miRNAs and designs new small RNAs
miRstring decodes mature miRNAs and designs artificial small RNAs across species, using an RNA language model. Small-RNA design shifts off hand-built motif rules and onto learned sequence representations, with cross-species transfer as the capability that matters most.
New method fuses models for gene-network inference
Fusing multiple models with sequence information improves gene regulatory network inference from single-cell transcriptomics, a bioRxiv preprint reports. Network inference remains among the least reliable steps in single-cell analysis, so accuracy gained there propagates into everything built on top of the networks.
Anthropic put Claude agents in a market
Anthropic built a market of Claudes to watch what works and what breaks when agents trade on people's behalf. Early evidence on delegated negotiation lands ahead of agents touching procurement and contracting in research operations.
Reply with your discoveries. A human reads them. Forward freely.
|