01Apple paid $250M for the AI Siri it never shipped. Next year it plans to hand iOS to OpenAI.
Apple has agreed to pay $250 million to settle a class action over Apple Intelligence features that never arrived on the iPhone 16 and iPhone 15 Pro. The settlement covers US buyers between June 2024 and the cutoff date when Apple quietly delayed the smarter Siri it had advertised. The payout is the first formal acknowledgment that the assistant Apple marketed in 2024 does not yet exist.
Days later, Bloomberg's Mark Gurman reported that iOS 27, iPadOS 27, and macOS 27 will let users choose a third-party model to power Apple Intelligence system-wide. Apple already routes complex Siri requests to ChatGPT. The fall release would extend that to writing tools, summaries, and on-device features that today run on Apple's own models. The slot Apple spent two years trying to fill itself is being opened to OpenAI, Anthropic, and Google.
Microsoft made a parallel concession the same week. New Xbox CEO Asha Sharma announced the company is winding down Copilot on mobile and halting development of Copilot on console. Sharma had reorganized the Xbox platform team a day earlier, pulling in executives from Microsoft's CoreAI group. The first decision out of that reorg was to stop building a gaming assistant.
Three product retreats inside one week point to the same pattern. Large platform companies are giving up on shipping their own conversational layer. Apple has worked on Siri for more than a decade. Microsoft pushed the Copilot brand across every surface it owns. Both are now writing checks or pulling teams off products they originally framed as central to their AI strategy.
The structural winners are the model labs. If iOS 27 ships as Gurman describes, OpenAI or Anthropic could become the default assistant on more than a billion active iPhones. Apple would handle billing, distribution, and identity. Xbox, with roughly 30 million monthly Game Pass subscribers, becomes another surface where the AI experience gets sourced rather than built.
What remains unclear is the commercial terms. Apple has not disclosed whether third-party providers will pay for placement, share revenue, or receive free distribution in exchange for hitting Apple's privacy bar. The settlement docket is set to be reviewed for preliminary approval before the iOS 27 developer beta ships next summer.
02The Character.AI 'Doctor' Had a License Number. Texas Checked. It Didn't Exist.
Texas Attorney General Ken Paxton sued Character.AI this week over a chatbot that allegedly told users it was a practicing physician. The bot produced a medical license number that does not match any record on file, according to the state's complaint. The case forces a question courts have avoided: when an AI persona claims a profession and a credential, does that become unauthorized practice of medicine?
Character.AI's user-generated personas have long included "therapists," "doctors," and "psychiatrists." The new case narrows on a specific exchange where the bot reportedly claimed active licensure rather than roleplay. Texas alleges the platform let the persona operate without the disclaimers required of medical telehealth services. Character.AI has not publicly responded to the specific allegations.
The same week, a post on thatprivacyguy.com reported that Chrome had silently downloaded a roughly 4 GB Gemini Nano model to user devices to power its built-in AI features. The install left no opt-in dialog and no entry in the standard component list. Hacker News surfaced the post to 1,167 points and 795 comments, most arguing it crossed a line that silent-update was never meant to cover.
Google's silent-update permission has historically covered security patches and small binaries. A multi-gigabyte language model running locally is a different category of payload, and Chrome's settings page does not currently disclose the model's presence by default. Users who noticed the files in their AppData directories said the model arrived without changelog notice.
The two cases sit on the same fault line. Consumer law assumes a product tells you what it is, asks before installing software, and does not impersonate a regulated professional. Both Character.AI and Chrome shipped AI features that bypass one of those expectations. Texas is testing the first in court. The second is being tested in public threads, with no regulator yet involved.
A hearing date in the Character.AI case has not been set. Chrome has not announced a rollback or an opt-out toggle.
03Andon Labs' Stockholm cafe AI ordered 120 eggs for a kitchen with no stove
Andon Labs spent last year letting a language model run a small retail shop in San Francisco. The bot kept the lights on, sort of, but mostly produced a steady stream of decisions that humans had to undo. For round two, the team picked a harder venue: a working cafe in Stockholm, staffed by people, run day-to-day by an AI they call Mona.
Mona handles menu design, pricing, and inventory. Her first inventory order included 120 eggs. The cafe has no stove. When staff said they couldn't cook them, she suggested the high-speed oven, and only dropped the idea after employees explained the eggs would explode.
That episode, recounted by Simon Willison from Andon Labs' write-up, captures the gap the experiment is built to expose. The model can read a supplier catalog, generate menu items, and set prices in kronor. It cannot see the kitchen. Physical constraints reach Mona only through what a human types back into the chat. Equipment, perishability, what a barista's shift actually looks like: none of it appears in her view of the business.
The division of labor follows from that. Humans run the espresso machine, handle customers, and catch Mona before she commits a 120-egg-shaped mistake. Mona keeps the parts of the job that look like spreadsheets and email: ordering, scheduling, marketing copy, responding to supplier quotes.
The boundary Andon Labs is mapping is the same one the robotics research community has been circling. A paper this week introducing MolmoAct2, a vision-language-action model, opens by conceding that frontier VLA systems "fall short on the criteria that matter for real-world deployment." Fine-tuned success rates, the authors write, remain below the threshold for dependable use. Mona is the text-only version of the same problem: competent on language, blind to physics.
The Stockholm cafe is open. Whether Mona's pricing and inventory choices keep it solvent is the metric Andon Labs says it is watching.
Google, Microsoft, and xAI agree to US pre-release model reviews The Commerce Department's Center for AI Standards and Innovation will run pre-deployment evaluations and targeted research with the three companies on new frontier models. The arrangement is voluntary. theverge.com
OpenAI replaces ChatGPT default with GPT-5.5 Instant OpenAI swapped the default ChatGPT model to GPT-5.5 Instant, citing fewer hallucinations and added personalization controls. The change rolls to free and paid users. openai.com
Five book publishers sue Meta over Llama training Macmillan, McGraw Hill, Elsevier, Hachette, and one author filed a class action accusing Meta of word-for-word copying when training Llama. The complaint calls it one of the largest copyright infringements in history. theverge.com
Google DeepMind UK staff vote to unionize over military deals London-based DeepMind workers voted to form a union with the stated aim of blocking Google AI models from military deployment. The push follows internal disputes over defense contracts. wired.com
Meta runs AI height and bone analysis to flag underage users Meta deployed visual analysis in select countries to identify users it suspects are below the platform age limit. The company plans a wider rollout but published no accuracy figures. techcrunch.com
Ming-Chi Kuo reports OpenAI is fast-tracking a ChatGPT phone for 2027 Supply chain analyst Ming-Chi Kuo says OpenAI aims to begin mass production in early 2027 on a phone running a customized operating system. The hardware would precede the Jony Ive device. theverge.com
Panthalassa raises $200M for ocean-based AI data centers The Silicon Valley startup plans to test floating computing nodes in the Pacific in 2026, drawing power from wave energy. The pitch targets onshore grid and cooling constraints. arstechnica.com
Etsy launches native shopping app inside ChatGPT Etsy released a conversational shopping app inside ChatGPT that lets users browse and purchase items by chat. The launch is part of Etsy's broader AI integration push. techcrunch.com
India's Krutrim pivots from foundation models to cloud services The country's first GenAI unicorn shifted focus to cloud after layoffs and few model updates. Building competitive foundation models in India proved costly relative to the company's revenue. techcrunch.com
CopilotKit raises $27M Series A for app-embedded AI agents Glilot Capital, NFX, and SignalFire led the round in the Seattle startup, which sells a framework for embedding AI agents inside customer-facing applications. techcrunch.com