Unreleased AI Models: Government's Interest

The Harvard Medical School study released this week puts OpenAI’s O1 model ahead of seasoned emergency‑room physicians in triage diagnosis. In 76 real ER cases the model achieved a 67 % exact‑or‑near‑match rate, compared with 55 % and 50 % for two attending doctors. Accuracy rose to 81.6 % after admission, edging out clinicians at 78.9 % and 69.7 %. While the results highlight AI’s diagnostic potential, the authors stress the test used only text inputs and warn that liability frameworks remain undefined when AI recommendations clash with physician judgment.

IBM’s Think 2026 keynote re‑positions Watson X as an AI governance and orchestration platform, allowing Fortune‑500 firms to deploy agents from Anthropic, OpenAI, Google and internal models under a single policy and audit layer. At the same time, PayPal announced a $1.5 billion cost‑cutting plan that leans heavily on AI, laying off 4,500 staff—about 20 % of its workforce—to fund an “AI turnaround.” Across the Atlantic, DeepMind’s London office voted 98 % to unionize, demanding the company withdraw from Pentagon and Israeli defense contracts, underscoring growing labor pressure on frontier AI labs.

The regulatory tide is rising as the U.S. Commerce Department’s Center for AI Standards and Innovation (CAISI) will now receive every new frontier model for national‑security red‑team testing before public release. This move follows Anthropic’s high‑profile Wall Street briefing, where CEO Dario Amadei and JPMorgan Chase’s Jamie Dimon showcased ten pre‑built finance agents designed to automate labor‑intensive workflows. Meanwhile, Cerebras’ $26.6 billion IPO, backed by a $1 billion loan from OpenAI, illustrates the massive capital flowing into AI chip makers and OpenAI’s strategic diversification away from sole reliance on NVIDIA. Together, these developments signal a shift toward tighter AI oversight, enterprise‑grade governance, and competitive supplier ecosystems.