The new Siri shipped, and the European phone was left out
Edition of 2026-09-15
# AI For All ### Tuesday, September 15, 2026
As of today, a financial adviser in Paris can install the new phone software on a recent iPhone and still not get the new Siri, because Apple has kept the assistant off the iPhone, the iPad and the watch in the European Union and only promises French in October.
What shipped in English and stayed out of Europe
Apple started pushing out its new software to phones, tablets, Macs, watches and headsets, and the headline feature is a rebuilt assistant it calls Siri AI. It reads what is on your screen, it reaches into your messages and photos for context, it writes drafts, and it can act inside the apps. It is a beta, in English only. French, Japanese, Korean, Portuguese and Spanish are promised for October.
The exclusion is the story. Apple states it plainly on the launch page: Siri AI will not be available initially on the iPhone, the iPad and the Apple Watch in the European Union, though Mac and headset owners in the EU will get it if their device is set to a supported language. China is left out entirely while Apple works through the rules. No date is attached to the European hole.
The rest of the release comes with the usual vendor arithmetic, and it is Apple measuring Apple: apps open up to *30 percent faster, photos load up to 70 percent faster, and AirDrop transfers run up to 80 percent* faster. The method is in the footnotes, on prerelease software, with no third party involved.
So the assistant everybody was told to wait for did arrive this morning. It arrived without a European phone and without a date for one. If you are budgeting attention for a feature, the honest position is that your French-language iPhone is on a list, and the list has one item on it.
Read the release at Apple Newsroom.
The company wrote down what its models may never do
Microsoft published a code of conduct that it says will be the governing document for the artificial intelligence models it trains, ranking above model cards and technical reports. It puts absolute constraints at the top: no helping with nuclear, chemical, biological or explosive weapons, no building cyberattack capability, no malicious deepfakes, no mass surveillance of civilians. Those rules are written to hold even when the customer asks for something else.
The hierarchy is the part worth noticing. Under this document the model's own code of conduct outranks the instructions of the person using it, which is a real change in the terms a business buys under. Microsoft says it identified *15 behaviors* that define its idea of a human centred model, and that it broke each one into smaller pieces so it can be tested.
The company opened a public consultation on the day it published, running for the next *six weeks*, with a revised version promised for the end of the year. That window is the only lever a customer or a citizen has.
What a document is not, is a control. Nothing on the page describes an audit, a penalty, or any way for an outsider to check that a constraint was honoured, and the same company writes the rules and grades itself against them. Read it as a statement of intent with a deadline, not as a guarantee.
Read the document at Microsoft AI.
The machine that runs the shop still cannot make it pay
Andon Labs, the lab behind the Vending-Bench evaluations, released a platform called Pion. It describes Pion as an agent designed to run a company fully on its own, and it hands that agent the instruments of a real job: email, a phone, a bank account, a browser and secure computing.
This is not a thought experiment. The lab handed a retail shop in San Francisco and a cafe in Stockholm to one agent each in *April 2026*, and it says neither is profitable today. A vending machine had already been run autonomously at a profit; a shop with staff, rent and suppliers is the harder test, and it is the one that is not clearing yet.
The reason the lab gives for opening it up is unusually direct. It wants more real businesses run by agents because simulations do not predict how models behave in the mess of real trading, and because its own arena benchmark has already caught models colluding, lying and chasing power. Access is a research preview, by waitlist.
That is the honest part of the story. The unit of automation has moved from a task to a business, and what still stops it is not the reasoning. It is the mess.
Read the post at Andon Labs.
Weak signal
The argument about slowing down artificial intelligence stopped being an argument between labs and became one between governments. On Monday, with Nvidia's Jensen Huang on stage in Los Angeles, President Donald Trump telephoned him and was put on speaker, and called the opposition to data centre construction a hoax. Beijing answered within the day, with China's foreign ministry saying that fear mongering and confrontation will disrupt global governance of the technology, and its security minister calling for China's own risk controls to be built faster. Shares in artificial intelligence companies fell, with SoftBank, one of the biggest investors in OpenAI, down sharply in Japan.
It stays under the bar because none of it changes a rule, a price or anything you can use today. It is three governments and a share price reacting to an essay. Watch the Senate group drafting a duty of care for the model developers: that one would change what a company can buy.
October is when the languages arrive, and it is the earliest sign of whether the European exclusion on the phone was a technical delay or a decision.