Medical startup uses LLMs to run appointments and make diagnoses (www.technologyreview.com)

Southern California startup Akido Labs is piloting ScopeAI, a proprietary LLM-based system that drives patient visits: medical assistants use ScopeAI’s prompts to interview patients while the system transcribes and analyzes dialogue, generates follow‑up questions, lists likely and alternative diagnoses, provides justifications and recommends treatment steps. The workflow uses fine‑tuned versions of Meta’s open Llama models and some Anthropic Claude models; a human physician then reviews and must approve or correct ScopeAI’s outputs before finalizing diagnoses or prescriptions. Akido says the approach lets doctors see four to five times more patients and has sped care for hard-to-reach populations—its street medicine team reports getting patients medications for opioid use disorder within 24 hours—while meeting Akido’s internal performance bar of including the correct diagnosis in its top three recommendations ≥92% on historical tests. The rollout is significant because it pushes LLMs from passive documentation or decision‑support into doing much of the cognitive work of a visit, raising regulatory, safety and equity questions. Legal and ethical issues include whether asynchronous AI-enabled care requires FDA clearance or runs afoul of medical‑practice rules, limited patient disclosure about AI involvement, insurer-driven disparities (Medicaid allows asynchronous approvals that some private plans don’t), and the risk of automation bias since physicians routinely defer to AI recommendations. Akido retrains models on clinician corrections, but outside observers call for randomized comparisons to standard care and outcome studies before broad adoption.