OpenAI rolls out Health in ChatGPT to US users
OpenAI is rolling out Health in ChatGPT to US users 18 and older across its Free, Go, Plus, and Pro tiers, on web and iOS. Users can connect Apple Health data and medical records from supported US hospital systems and providers like One Medical, then ask ChatGPT to compare lab results over time or explain what changed since a visit. The feature first appeared in January for a small tester group; this is the rebuilt version now reaching the general public. OpenAI says the connected data won’t train its foundation models or target ads, and gets deleted within 30 days of disconnecting. What deserves scrutiny is the shift in the trust structure: once medical records move from healthcare institutions into a consumer app — a setting that HHS guidance says generally falls outside HIPAA unless the app is acting on a provider’s behalf — their use is constrained mainly by OpenAI’s own privacy promises and general consumer-protection law. Health data flowing into a general-purpose chatbot at consumer scale makes its privacy and liability boundaries a defining Responsible AI question for the coming year.
Nearly 200 startups urge Washington not to ban Chinese open-weight models
Nearly 200 startups — most US-based, including Y Combinator — formed the “Little Tech Association” and sent letters to the White House and Commerce Department opposing restrictions on Chinese open-weight models — models whose weights are freely downloadable, such as Moonshot’s Kimi and Alibaba’s Qwen. Their core argument: a ban wouldn’t stop these models from spreading globally, but it would cripple the US startups that build on them; targeted safeguards should replace blanket prohibitions. This continues the sanctions saga from earlier this week, and adds a new element: the coalition’s first coordinated pushback in this debate. The same day, TechCrunch quoted experts disputing the official claim — leveled by White House science adviser Michael Kratsios — that Kimi K3 was built by distilling Anthropic’s Fable, distillation meaning training your model on another model’s outputs. Braden Hancock of the Laude Institute notes Fable only became publicly available July 1, making a distillation-driven K3 implausible on the timeline; Ai2’s Nathan Lambert argues frontier-level gains come mainly from reinforcement learning, which distillation through an API can’t deliver cheaply or effectively (see also TechCrunch). The accusation has come with no public evidence so far, while the rebuttals rest at least on verifiable release dates — and if policy gets built on an unverified technical narrative, those 200 companies pay the price first.
Nikkei investigation: five tech giants carry $1.65 trillion in off-balance-sheet AI debt
Nikkei Asia estimates that Microsoft, Alphabet, Amazon, Meta, and Oracle carry roughly $1.65 trillion in off-balance-sheet obligations — long-term data center leases, GPU and server supply contracts, and other financing arrangements — more than their combined on-book debt of about $1.35 trillion, with Meta alone near $420 billion. To be precise: these arrangements are GAAP-compliant, not Enron-style fraud, but the effect rhymes — the real leverage behind AI infrastructure doesn’t fully show up on the balance sheets. The mechanism that matters is commitment without visibility: the long-term contracts bind, meaning the companies owe those payments even if AI monetization disappoints, yet the scale of the leverage — and who ultimately bears it — is hard to read off any balance sheet.
DARPA and the US Air Force fly an AI-controlled operational F-16
DARPA announced that its VENOM program completed test flights in June at Eglin Air Force Base with an AI agent flying a standard operational F-16; a human pilot stayed in the cockpit throughout, able to toggle between manual and autonomous control. An AI flying a fighter jet isn’t new — the X-62A VISTA testbed ran AI dogfights under DARPA’s ACE program in 2023–2024. The increment here is industrialization: instead of a one-of-a-kind testbed, VENOM retrofits standard fleet aircraft with an autonomy kit without touching core software. Military autonomy is crossing from “can it fly” to “can it be rolled out across a fleet” — and the governance conversation needs to move at the same speed.
Gemini CLI patches a remote code execution vulnerability
The latest Gemini CLI nightly fixes a vulnerability in its a2a-server — the server component of the agent-to-agent protocol — where workspace trust and task isolation weren’t enforced, opening the door to remote code execution. The changelog is a single line pointing to the fix, and the release notes list no CVE. This one earns its own entry because it confirms what recent agent security research keeps warning: a coding agent holds execution rights on your machine, and every protocol component it ships is attack surface. Teams using these tools should update promptly and review their workspace trust settings.
Gemini closes in on a billion monthly users
Google disclosed on its Q2 earnings call that Gemini now has over 950 million monthly users, up from 750 million in February. ChatGPT reached a billion monthly active users in May, per Sensor Tower — but the same report shows its share of AI-assistant usage, measured as unique users across mobile apps and web, fell below 50% for the first time in March. A consumer AI duopoly is taking shape — and distribution through Android, Search, and Workspace is exactly the lever Google is pulling.
Ethan Mollick updates his “which AI for what” summer guide
Wharton professor Ethan Mollick published his summer 2026 edition of which model to use for which task — an avowedly opinionated guide drawn from his own use of the models. Official documentation rarely keeps pace with how fast models change, and in my view a firsthand guide like this is one of the more useful references an ordinary user has. Worth bookmarking.
One-line takeaway: Medical records flowing into a chatbot, a fight over banning open weights, $1.65 trillion off the books — today’s top stories all point the same way: the next phase of AI competition is less about model capability than about trust structures, and whoever can answer “why should anyone trust you” wins the next round.