> daily_signal(2026_09_10)
Microsoft barred AI from training on kids' school data, Apple put always-on listening on your wrist, JD.com ordered 3 million robots to cut 700,000 jobs, and an AI learned to pick a tumor's drugs.
PickBits Daily Signal · Thursday, September 10, 2026
// tl;dr
- Microsoft and the two biggest teachers' unions signed a 30-page, enforceable standard that bars AI companies from training on students' data. A district can add it to its Microsoft contract with no renegotiation from November 1, with 72-hour breach reporting, annual audits, and the right to cancel and sue. It binds Microsoft only; OpenAI and Anthropic have not signed.
- Apple put always-on listening on the Apple Watch and published an 11-page paper swearing the microphone audio never leaves a sealed chip. Every feature is opt-in and runs in a "Secure Exclave" on the new S11 chip that Apple says even it can't reach. The catch: the whole privacy case rests on a hardware claim buyers cannot check.
- JD.com said it will buy 3 million robots over five years, and its founder said they will eventually replace all 700,000 of the company's delivery and warehouse workers. A "Nirvana Plan" promises to retrain couriers into robot-repair roles. It's the most concrete robots-for-jobs number yet, though a five-year buying target is a plan, not a delivered fleet.
- An AI "virtual cell" read tumors' proteins and predicted drug response in the hardest breast cancer with 88% accuracy on 81 drugs it had never seen. Its predictions matched what actually happened to 501 real patients. It's a research model that still has to pass clinical trials before it reaches a real patient.
Two of these we've been following for weeks. The fight over AI in schools has been all government so far: Florida moving to force every district to adopt a policy, New York City pulling generative AI out of its youngest classrooms days later. This week the vendor side finally answered, and it put the answer in a contract. And the jobs-versus-robots number we first saw in June, when Amazon's leaked plan projected replacing hundreds of thousands of jobs, came back today with China's JD.com saying the quiet part out loud: three million robots, and a founder who says he will replace all 700,000 of his couriers.
The other two are simpler. Apple put a microphone that's always ready to listen on your wrist, then published eleven pages arguing a sealed chip is why that's fine, which it might be, and which you can't check. And the day closes on the one I would actually want to win: an AI that reads a tumor's proteins and ranks the drugs most likely to work on the breast cancer with the fewest options. My own read is that on the two stories aimed straight at you, the school data and the wrist microphone, a written promise you can audit is worth more than any demo.
Microsoft will let a district sue it over student data; Apple just asks you to trust a chip you can't open.
1. Microsoft signed a binding standard barring AI training on students' data.
It becomes enforceable inside a district's Microsoft contract on November 1, with audit rights and the right to sue, and it binds only Microsoft.
For a year the pressure on AI in schools came from governments. Two weeks ago Florida moved to make every district adopt an AI policy; New York City pulled generative AI from its K-8 classrooms days later. On September 9, the biggest ed-tech vendor answered. Microsoft, standing with the American Federation of Teachers and the United Federation of Teachers, published a 30-page National AI Safety & Privacy Standard that a district can bolt onto its existing Microsoft agreement with no renegotiation, effective November 1. Vice Chair and President Brad Smith summed up its ten protections in three words: privacy, safety, transparency.
The terms are the story. Student and educator data cannot be used to train AI models or sold for advertising. Student tracking and AI companion chatbots aimed at kids are prohibited. A human has to review any consequential AI decision, breaches must be reported within 72 hours, and Microsoft submits to annual certification and mandatory audit rights. Break any of it and a district can cancel the contract and sue for damages. It binds Microsoft, though, and only Microsoft. The unions said they hope OpenAI and Anthropic follow; neither commented. And a contract term is only as strong as a district's willingness to enforce it.
Why this matters: A major AI company just agreed in writing, and under threat of a lawsuit, never to train its models on your kid's school data. A policy statement, a district can ignore. This one it can audit, cancel, and sue over, and it answers the classroom bans cities reached for two weeks ago. The catch worth keeping in view is that it binds Microsoft alone, so every other AI vendor now gets measured against it.
Action this week: Read the actual Memorandum of Agreement before November 1 and confirm with your Microsoft account team which of your tenants and products the ten protections actually cover, because they are opt-in to your agreement, not automatic. When I have pushed a vendor on data use, the useful answer is always the specific one, so get it in writing rather than in a slide. And if you are a parent, ask your district two questions this month: is it adding the standard, and what plain-language guide will it give families about how these tools use student data.
fortune.com: Microsoft agrees to an enforceable AI safety and privacy standard for school districts (September 9, 2026)
news.microsoft.com: AFT, UFT and Microsoft announce a National AI Safety & Privacy Standard for schools (September 9, 2026)
edweek.org: Microsoft agrees to new student-privacy protections for AI, how ironclad are they (September 2026)
2. Apple put always-on listening on your wrist.
It published eleven pages swearing the microphone audio never leaves a sealed chip, and the whole privacy case rests on a claim you cannot check.
There's no independent test you can run on this one. On September 9, Apple shipped Audio Intelligence on the Apple Watch Series 12 and Ultra 4 — Sound Recognition and automatic Shazam music recognition now, with a Live Rewind that keeps the last 15 seconds of a conversation and a Siri Recap both arriving in beta later this year — then published an 11-page paper arguing the features can listen without becoming surveillance. All of it runs on a Secure Exclave inside the new S11 chip. Apple says the raw microphone audio flows into a buffer there, is never written to a file, and is inaccessible to the operating system, to apps, and even to Apple itself.
Every feature is opt-in. Live Rewind fires only when you deliberately double-press the Digital Crown, keeps a short on-screen text snippet rather than a recording, does not transcribe in the background, and chimes and shows an indicator so people nearby know. Anything you save syncs end-to-end encrypted with Apple holding no keys. On paper that's stronger than a cloud voice assistant. It is also unverifiable from the outside, described in a document Apple wrote about its own product. Keep TechCrunch's read next to Apple's: a microphone that is always ready to listen is quietly becoming a normal thing to wear.
Why this matters: Your new Apple Watch can be always ready to listen, and the only reason that's supposed to be fine is a chip you will never be able to open and check. The architecture Apple describes really is stronger than sending your voice to the cloud, and it's also unverifiable from the outside, and a microphone you wear all day is something you will just stop noticing. The whole privacy case is a hardware claim, and the only honest thing to do with a hardware claim is go and check it.
Action this week: Open the Watch app or Control Center and switch on only the listening features you actually want, and leave the rest off, because always ready to listen is a setting, not a default you're stuck with. If you're ever near someone using Live Rewind, the chime and indicator mean their watch just kept fifteen seconds, and you can ask about it. And if you work in security or policy, read the Secure Exclave paper and then push for someone independent to test it, because the whole privacy claim depends on a chip boundary buyers cannot check for themselves.
9to5mac.com: Apple details how the Apple Watch's new Audio Intelligence features work in a privacy paper (September 9, 2026)
techcrunch.com: The Apple Watch's new AI features are normalizing the idea that technology is always listening (September 9, 2026)
apple.com: Audio Intelligence Privacy Overview (September 2026)
PickBits Daily Signal is free. If it lands in your inbox every day and it is worth something to you, the best way to support it is to forward it to someone who would read it. Subscribe today!
3. JD.com ordered 3 million robots to automate its warehouses and deliveries.
Its founder said the robots will eventually replace all 700,000 of the company's couriers, with a retraining program as the cushion.
If you run or plan a supply chain, this is the number to put on the wall: three million robots. At its Global Technology Explorers Conference on September 9, China's JD.com announced a Physical AI Acceleration Plan to buy 3 million robots, 1 million autonomous vehicles, and 100,000 drones over five years, built around a new industrial Wolf Robot line for picking, sorting, transport, and delivery. Founder Liu Qiangdong didn't soften it: the robots will eventually replace all 700,000 delivery and logistics workers JD and its ecosystem employ.
We covered the American version of this in June, when Amazon's leaked plan projected replacing hundreds of thousands of jobs with robots. JD is the same move, said out loud and put on a five-year calendar. The cushion is a Nirvana Plan to retrain couriers into robot servicing and maintenance, which is exactly the promise to watch, because every automating employer is going to make it. A five-year procurement target is a plan, not a delivered fleet, and a retraining pledge is easy to announce and hard to audit. It's a China story, but the same robots-versus-jobs math is heading for supply chains here too.
Why this matters: A company just put a real number on replacing you with a robot: three million machines to do the picking, driving, and delivering that 700,000 people do today, said out loud by its own founder. This is the most concrete robots-for-jobs commitment yet, and the retraining pledge is the part to hold them to, because "we will retrain the people we replace" is the line every employer will reach for. So watch one number, and it is not the robot count: how many of those 700,000 workers actually land in the new jobs.
Action this week: Track deployment numbers rather than announcements, and ask which specific tasks the Wolf Robot line actually does reliably today and at what cost per task versus a human, because a five-year target tells you intent, not capability. If you shape labor policy, get the real figure in writing: not how many robots a company buys, but how many displaced workers it moves into the new technical roles. And if your own job sits in picking, sorting, or driving, start now on the adjacent skills that are harder to hand a machine, the ones JD's own retraining plan is betting on.
scmp.com: JD.com to deploy 3 million robots to fully automate its logistics (September 9, 2026)
sourcingjournal (wwd.com): JD.com founder says robots will replace 700,000 couriers, with retraining into repair roles (September 9, 2026)
digitimes.com: JD.com expands logistics robotics with its Physical AI Acceleration Plan (September 9, 2026)
4. An AI virtual cell learned to pick the hardest breast cancer's drugs.
It hit 88% accuracy on 81 drugs it had never seen, and its predictions matched what actually happened to 501 real patients.
Picture someone just diagnosed with triple-negative breast cancer, the aggressive subtype with the fewest targeted treatments, where the current answer is too often to try one drug, wait, and try another while time runs out. A study published in Nature on September 9 (Sun et al.) describes an AI virtual cell that reads a tumor's protein makeup and predicts how its cells will respond to a drug. Trained on millions of protein measurements, it reached 88% accuracy predicting the response to 81 drugs that were not in its training set, which is the hard test of whether a model learned real biology or memorized its examples.
Then the researchers ran it on 3,651 proteins measured in pre-chemotherapy biopsies from 501 patients, and its predictions tracked the outcomes those patients actually went on to have. Along the way it flagged the specific proteins that drive drug resistance, the molecular reason a treatment fails. One caveat, and it is a real one: this is a research model whose value still has to be proven in prospective clinical trials, and an 88% in-silico score is a reason to run the trial, not a treatment decision. AI tools for breast cancer have fallen short of radiologists before, so this one earns trust in a trial, not a demo.
Why this matters: An AI learned to read a tumor and rank which drugs will beat the hardest breast cancer, and it was right 88% of the time on drugs it had never seen. The payoff, if trials bear it out, is less trial-and-error for patients who have the least time to spare, and a tool that helps an oncologist choose rather than a chatbot that pretends to diagnose. Keep the caveat attached: a prediction that matched 501 real outcomes is a reason to test this in people, not a test you can ask for yet.
Action this week: Ask your oncologist or a molecular tumor board whether proteomic or molecular profiling is available for your case, and whether a trial near you is using predictive tools to guide drug choice, while knowing these models are not yet standard of care. What I will be watching is whether the proteins it flags for resistance hold up as real biomarkers in other patient groups, because that is the line between a striking result and a clinical tool. And if you fund science, this is exactly the kind of problem the market skips: precision treatment for an aggressive, under-served cancer.
nature.com: An AI virtual cell predicts drug response in triple-negative breast cancer (September 9, 2026)
nature.com: Sun, R. et al., proteome-based virtual cell model of drug response (September 2026)
» What to watch this week
- Does any other AI company sign something like Microsoft's school standard. It only sets a real floor if OpenAI, Anthropic, and the districts themselves take it up before November 1.
- Whether anyone outside Apple actually tests the Secure Exclave. The only proof the audio stays on the chip is Apple's own paper, and an outside audit is the one thing that could change that.
- JD's real reskilling count, not its robot count. The Nirvana Plan is just a promise until the company reports how many of the 700,000 moved into repair jobs instead of out the door.
- A registered clinical trial for the virtual cell. An 88% in-silico score is a reason to run one; the milestone that counts is a study testing whether its drug picks help real patients.
Tomorrow's signal lands here.