FDA floats doctor-style licensing exams for generative AI medical devices

Share
FDA floats doctor-style licensing exams for generative AI medical devices

The FDA is weighing whether generative AI health products should pass "competency-based" exams — modeled on how human clinicians are trained, tested, and credentialed — before they can be marketed as medical devices. The idea sits in a discussion draft the agency's device center published this week, and it is explicitly a question, not a rule: the FDA says the paper is "for discussion purposes only," with public comments open until October 19.

The proposed framework would test the final, user-facing product rather than the underlying foundation models — checking whether an AI device demonstrates clinical knowledge, analytic capability, safety behavior, communication skill, and generalizability for its intended use. Test rigor would scale with risk: a product that nudges a user toward a specific action would face a harder bar than one that only supplies information, and testing alone might not suffice — the agency is also weighing heavier reliance on real-world evidence after devices reach the market.

It is a genuinely new regulatory shape. Licensing-style testing has circulated in academia for a year, and the American Medical Association is mulling whether AI licenses make sense — but this is the first time the FDA itself has put the concept into an official public document. Generative AI has been notoriously hard to fit into device regulation because models keep evolving and their outputs vary run to run; a competency-exam-plus-postmarket-surveillance model accepts some premarket uncertainty in exchange for ongoing scrutiny.

Our take: the exam framing gets the hard part right by targeting what ships to patients rather than the model underneath, but it also imports medicine's weakest ritual — the boards are a floor, not a guarantee of bedside judgment. If the FDA finalizes anything along these lines, expect it to become the template other regulators copy, which is exactly what the agency says it wants.

What to watch: the comment docket closing October 19 — and whether any draft guidance keeps the clinician-style exam or retreats to conventional performance testing.


SoftBank plans a record $6.3 billion retail bond sale to fund its AI bets

SoftBank is preparing a roughly ¥1 trillion ($6.3 billion) bond sale aimed at retail investors — the largest retail-targeted corporate bond ever issued in Japan, and its third such offering this year alone. Bloomberg reported the plan over the weekend, following Nikkei's earlier disclosure of the offering's size; SoftBank set its previous retail record last year at ¥600 billion, so this more than doubles it.

The money feeds Masayoshi Son's OpenAI machine. SoftBank plans to pour around ¥10 trillion into OpenAI-related commitments by October, and Nikkei flagged that proceeds could go toward repaying the bridge loan — arranged through Goldman Sachs, JPMorgan, and Apollo against SoftBank's existing OpenAI stake — that funded the additional investment. Proceeds are also earmarked for debt repayment and physical-AI acquisitions such as ABB's robotics division and Swiss construction-robotics startup Gravis.

Two details deserve attention. First, the yield on the coming seven-year bonds is expected to sit in the high 4% range — well above the 3.15% SoftBank paid in late 2024 — which tells you the market is charging Son more for leverage every quarter. Second, Japan's retail bond market is being drained by AI borrowers: total retail corporate issuance this year is projected at a record ¥2.8 trillion, with SoftBank the biggest single name. Households' savings accounts are now quietly underwriting the AI buildout.

Our take: this is Son monetizing his own popularity in Japan to keep pace in a race he cannot afford to slow down in — a clever channel while Arm's share price props up his balance sheet, but each successive record issue makes the whole structure more rate-sensitive.

What to watch: the coupon when the deal prices — if it clears 5%, that is the market repricing OpenAI leverage in real time.

Which experiment will matter more — the FDA teaching AI to sit for boards, or Tokyo households bankrolling AGI? Tell us in the comments.

Read more

OpenAI in early talks to buy security firm Trail of Bits

OpenAI in early talks to buy security firm Trail of Bits

OpenAI's security ambitions got a name attached to them today, Google opened its frontier models to European scientists, and a Chinese lab put an agent inside a phone and sold it by the token plan. OpenAI is in early talks to acquire Trail of Bits, the New York security firm that has spent fourteen years finding and fixing vulnerabilities in open-source software. The Information reported the discussions Tuesday, citing people familiar with the matter; the deal size is undisclosed, the talks ar

Deep Dive — Drones just proved AI compute is a wartime target

Deep Dive — Drones just proved AI compute is a wartime target

Four days of Ukrainian drone strikes took out half of Yandex's data centers, including the buildings that hold all three of its training supercomputers. The rest of the industry is now asking a question it had been avoiding: what happens when the machines that make AI are also the easiest thing to hit. The four days Russia started it. Through late September, Russian drones and missiles hit Ukrainian data centers and telecom infrastructure, temporarily cutting internet and cell service for u

Tencent slots an AI agent into 14 million companies' contact lists

Tencent slots an AI agent into 14 million companies' contact lists

China's biggest workplace chat app now lists an AI agent as a colleague. Tencent shipped a deep two-way integration between WeCom and its WorkBuddy office agent, and it went out to the version of WeChat that businesses actually run on. WorkBuddy now sits in the WeCom contact list as a conversational bot — not a sidebar plugin. In WeCom 5.0.12, the agent appears next to human colleagues and can be messaged like one; WorkBuddy 5.8.0 in turn logs in with a WeCom account and inherits permissions a

Drone strikes knock out three Yandex data centers in four days

Drone strikes knock out three Yandex data centers in four days

Russia moved its digital economy onto a handful of centralized campuses. Ukraine just demonstrated what that concentration costs — and two of the buildings hit hold the machines Yandex trains its AI on. Ukrainian drone strikes have knocked out three of Yandex's data centers in four days, including the sites that house the supercomputers the company uses to train YandexGPT. The first attack damaged the Sasovo facility southeast of Moscow on October 8. That campus holds two of Yandex's three AI