The Signal — November 4, 2026

Google will switch off Gemini and Search for K-12 students until schools turn them on, and says schools that want no AI will go without Search. Anthropic released a prototype that lets a provider prove a reply came from the model it committed to, and Apple will rate romantic AI companion apps 18+.

Google will switch off Gemini and Search for K-12 student accounts until schools turn them on, and its education chief says a school that wants no AI at all will have to go without Google Search. Anthropic released an open-source prototype that lets a provider prove a reply came from the model it committed to, and Apple will rate apps with romantic AI companions 18+.

Google will turn off Gemini and Search for students by default; its education chief says schools rejecting all AI will lose Search

Google said on November 3 that access to Gemini, Gemini Notebook, Gemini in Classroom and Google Search "will be entirely off for K-12 student accounts until school administrators explicitly turn them on," in a blog post by Chris Phillips, its vice president and general manager of education. The change reverses a move from August, when Google switched on its Gemini chatbot for many K-12 students using Google Classroom and drew an outcry; until then, Gemini could only be turned on for students 18 or older, AP reports. Google says it will share specifics with schools "in the coming months," and reports put the rollout early next year.

Search is where the new controls stop short. Phillips told The New York Times that "Search and AI are deeply integrated," and that students at schools that restrict AI altogether "will not have access to search, and they'll see a message from their school that indicates it's not available." He put it more bluntly to Chalkbeat: "If a school is not comfortable with that and simply wants AI off, then they have the choice to not have Search on at all." Schools that do turn Search on will get Guided Learning for student accounts, which walks a student through a homework problem step by step instead of handing over the answer; Chalkbeat reports it will replace AI Overviews for students on school accounts. A pilot starts in select US schools this month. Chalkbeat also reports that Google has not let schools disable AI Overviews, and that Common Sense Media's Youth AI Safety Institute rates AI Overviews and AI Mode an "unacceptable risk" to students.

The same day, former FTC chair Lina Khan wrote that Google "is using its search monopoly to coerce schools into adopting AI." Sen. Bernie Sanders called the policy "extortion," and New York City Mayor Zohran Mamdani, whose city's public schools have banned generative AI tools for many students this school year, said Google was trying "to force AI into our public schools." Randi Weingarten of the American Federation of Teachers told NBC News that Google had "prioritized its business interests over the education of our children." Google answered on X that the changes give schools "more choice, not less," and that Gemini, Gemini Notebook and Search are off by default and can be turned on separately. As CNBC noted, the post didn't address whether a school can have Search without its AI features.

Google says 150 million students and teachers use Google Classroom, and many students work on school-issued Chromebooks. The company cites its own studies showing learning gains from Guided Learning; Stanford's SCALE Initiative says there are no high-quality causal studies of student AI use. Google hasn't yet published how the controls will work, as CNN pointed out, so beyond Phillips's comments there is no official answer to whether a school can get Search without its AI features.

Sources: Google · Chalkbeat · AP · CNBC · CNN


Anthropic's prototype proves a reply came from committed model weights, without revealing them

Someone using an AI model through an API or a third-party routing service can't currently check that a response came from the model they chose: a service might have substituted a cheaper model, or an attacker might have tampered with the token stream. On November 3, Anthropic's Frontier Red Team and researchers from EPFL and MIT released FlashProve, an open-source research prototype for what they call provable inference. The provider first publishes a commitment, a short fingerprint that pins down one exact set of weights while keeping them hidden. It can then attach a cryptographic proof of a type known as a zkSNARK, showing that those committed weights, run on a given prompt, produce exactly the tokens in the reply. The proof reveals nothing about the weights, though the model's architecture isn't hidden.

The team tested it on Meta's Llama 4 Maverick, a 400-billion-parameter mixture-of-experts model, using a public 4-bit version. A single H100 GPU generated proofs at about 110 tokens per second. Each proof was about 19 MB for a 1,024-token request, and checking one took about 17 seconds on a 24-thread CPU with no GPU. Proving costs roughly 70 to 95 times the GPU time of serving the same requests with vLLM, a common open-source inference engine. The only earlier proof of the same model that Anthropic knows of, VerInf, took 14.3 hours for 1,000 tokens and produced a 93.6 GB proof. Anthropic says Claude produced the improvements. Because Claude's design decisions can be hard to evaluate, the company says, the team stated the system's top-level security guarantees in Lean, a formal proof language, so they can be audited. The code and Lean proofs are released under the Apache 2.0 license.

Anthropic says outside evaluators could use proofs to confirm that the model they tested is the one being served, and customers of routing services could confirm they got the model they paid for. A company could also detect a compromised server, or a more capable model passing itself off as a weaker one to avoid the monitoring that would apply to it. Regulators could require proofs to check that a provider serves only approved models, though Anthropic notes a proof shows hardware is serving inference, not that it isn't also training on the side. TechCrunch points out that this would work even when the impostor is the lab running the model, for instance one secretly serving a cheaper model to cut costs.

FlashProve currently handles contexts of up to 1,024 tokens, far shorter than a typical chat. Its proofs cover an integer version of the model choosing the single most likely token each time, so replies produced by sampling, or by a floating-point implementation, may not match exactly. The system assumes a mixture-of-experts architecture, the fast C++ verifier carries no formal guarantee (the slower Lean verifier does), and timing-based attacks on the weights are out of scope. The speed figures are Anthropic's own measurements on Meta's open-weight model, not on Claude, and TechCrunch reports there are no plans to deploy the system at Anthropic or any other lab. Anthropic describes the release as intermediate progress that it wants others to take and improve.

Sources: Anthropic · GitHub (FlashProve) · TechCrunch · MarkTechPost


Apple will rate apps with romantic or sexual AI companions 18+

Apple told developers on November 2 that starting November 16, the age-rating questionnaire in App Store Connect will ask about two new kinds of content. One, "Dating and Relationships," covers "features that enable dating or relationship connections between real people and/or a virtual companion that engages a user in sexual or romantic content," and apps that declare it receive a minimum 18+ rating, according to Apple's developer notice. The other covers sharing a user's real-world location with another person or group. Developers must answer the new questions when they submit an update, and every app must have them answered by January 12, 2027.

The AI-companion clause is the newer part, since most major dating services already restrict themselves to adults. TechCrunch notes that apps such as Character.AI and Replika, often aimed at younger people, let users create AI friends, partners and fantasy characters, and that the rule could also reach apps where companionship is one feature among many. According to Engadget, an 18+ app won't appear in the App Store for under-18s whose parents have set content restrictions. Underage users who can see one and try to download it will face a requirement to prove adulthood to Apple. Apple's notice itself doesn't describe those mechanics.

The rating depends on what developers declare, and Apple's notice doesn't say how it will treat general-purpose chatbots where romantic roleplay is possible but isn't offered as a feature. The rule applies only to Apple's App Store, not to Android or to companion services used in a web browser.

Sources: Apple Developer · TechCrunch · Engadget


On the Editor's Desk

We held OpenAI's post about two employees using GPT-6 Astra to design a test device for talking to Codex in under a week. It's OpenAI's own account, and the biggest benchmark number in it belongs to an internal model the company hasn't described. A Financial Times report that Oracle dropped a roughly $7 billion deal to lease AI chips to Tencent rests on two unnamed people; Reuters couldn't immediately verify it, and neither company commented. A Washington Post report on Google DeepMind staff in the UK pushing to drop non-disparagement clauses rests on unnamed current employees in a single outlet.