Will Google schools lose Search if they refuse Gemini?

For school-managed K-12 student accounts, Google says Gemini, Gemini Notebook, Gemini in Classroom, and Google Search will be off by default until administrators explicitly enable them, with rollout aimed at early 2027. Chris Phillips (Google education VP) told The New York Times that Search and AI are deeply integrated, so districts that block Google’s AI tools also lose student Search access. Opt-in schools choose either a consumer-style Search with AI Overviews/chat or a Guided Learning mode.

Does ChatGPT really have 20% of Google’s search market share?

No independent market-share audit was published with the claim. ChatGPT consumer product lead Adam Fry told Big Technology that ChatGPT’s volume of search-like queries is roughly one-fifth of Google’s query volume. OpenAI counts a search each time a user question triggers a web search — not every chat turn, and not referral traffic to publishers. Fry did not disclose the Google baseline, period, or geography, so treat it as an OpenAI estimate, not verified share.

Is Anthropic FlashProve deployed on Claude today?

No. FlashProve is an open-source research prototype for cryptographic provable inference (zkSNARK-style proofs that a reply came from a committed model without revealing weights). Anthropic’s Nov 2, 2026 research post says it is intermediate progress with no plan stated to deploy it on Anthropic or other labs’ production serving. Demo numbers use Llama 4 Maverick (400B MoE) on a single H100 at about 110 tokens/second of proving with ~19 MB proofs per 1,024-token request.

AI Daily · wordok.top

Wednesday, November 4, 2026

Google ties K-12 Search to an explicit AI opt-in, OpenAI floats a ChatGPT search-volume figure against Google that needs careful definition-reading, and Anthropic opens FlashProve — a zkSNARK-style path to prove which model answered without shipping weights. Three lanes for school IT, search marketers, and trust/security teams.

Published 2026-11-04 · Asia/Shanghai · Original briefing · not a digest reprint

Policy — Google makes K-12 Search ride with Gemini

On November 3, 2026, Google told K-12 administrators (via The New York Times, AP, The Verge, and Ars Technica) that for school-managed student accounts, Gemini, Gemini Notebook, Gemini in Classroom, and Google Search will be entirely off by default until admins explicitly turn them on. Rollout is aimed at early 2027, with more district-facing detail promised in the coming months. The move walks back the August default of Gemini inside Classroom for many schools that already allowed the chatbot.

The catch is the bundle. Education VP Chris Phillips said “Search and AI are deeply integrated”: districts that refuse Google’s AI stack also lose student Search, with students seeing a school message that the service is unavailable. Opt-in schools get two Search flavors — a consumer-style experience with AI Overviews / chat, or a Guided Learning mode meant to tutor step-by-step rather than dump finished answers. Guided Learning for schools is being piloted in select U.S. schools this month; Google says teachers can review chatbot transcripts in the school version. Lawmakers and city leaders (including Sen. Bernie Sanders and NYC’s mayor in contemporaneous coverage) framed the Search cut as coercive; Google frames it as more administrator choice after educator pushback.

Why it matters here: Chromebook-heavy districts cannot treat “AI off” as a soft content filter anymore — it is a research-tool outage unless they accept Gemini-class features or rebuild discovery on non-Google stacks. For product and policy readers on wordok.top, this is the education twin of consumer AI Overviews: search UX and generative answers are one control plane. Procurement checklists should ask whether “Search without generative AI” is even offered.

Markets — OpenAI’s “~20% of Google” search figure needs a ruler

At an OpenAI-hosted New York dinner on Monday night (Nov 2), ChatGPT consumer product lead Adam Fry told Big Technology (transcript posted Tuesday) that ChatGPT’s volume of search-like queries is roughly one-fifth of Google’s query volume. He also said weekly active users moved from ~1B to ~1.2B in about 60 days, with ~1.5B images created per week and ~250M combined weekly voice/dictation users. The Verge, Search Engine Journal, and others amplified the search comparison on Nov 3–4.

Definition is the story. Fry said OpenAI counts a search every time a user question triggers a web search — not every chat turn, and not the same unit as classic “ten blue links” sessions. He did not publish the Google baseline, the window, or geography. Outside trackers that measure referral clicks to publishers still show Google dwarfing ChatGPT outbound traffic; those are different meters. Bing’s long-cited ~3% search share (2023 trial testimony) is a share-of-engine frame, not the same as OpenAI’s internal search-trigger count.

Why it matters here: marketers and SEOs on wordok.top should split dashboards into (a) assistant search-trigger volume, (b) classic SERP share, and (c) referral sessions. Quoting “ChatGPT is 20% of Google” without the counting rule will mislead boards. Habit shift is real; the competitive map still needs two columns for query volume vs clicks sent to the open web.

Research — FlashProve: prove the model without shipping the weights

Anthropic’s Frontier Red Team published “Progress on provable inference for large language models” on November 2, 2026 (TechCrunch and others covering into Nov 3–4), releasing FlashProve: an open-source prototype that lets a provider commit to model weights (a short fingerprint that hides the weights) and later attach a zkSNARK-style proof that a given reply was produced by that committed model. Anyone with the commitment, prompt, and reply can verify on ordinary CPU hardware. Code, Lean formalization of top-level security claims, and docs are under Apache 2.0; Anthropic frames it as intermediate research, not a finished product roadmap for Claude serving.

Demo scale: on Llama 4 Maverick (400B MoE, public 4-bit weights), a single H100 generates proofs at about 110 tokens/second, with proof size ~19 MB per 1,024-token request — roughly 70–95× the GPU time of running the same batch on vLLM. Offline commitment takes ~39 minutes per model (good for on the order of ~16M requests before recommit). Stated use cases include spotting cheaper/swapped models behind routers, helping auditors verify evals, supporting pacing/compliance narratives, and provenance for “this library was scanned by model X” claims. Limits are explicit: greedy/integer decoding semantics, short 1,024-token context, MoE-oriented design, architecture not hidden, and no claim that fast C++ code matches the Lean proofs bit-for-bit.

Why it matters here: enterprise buyers already ask “which model answered?” when APIs sit behind gateways. FlashProve is not production Claude attestation today, but it moves cryptographic model identity from theory into a runnable artifact auditors can poke. Pair it with Anthropic’s separate September 2026 inference-verification prototype goal (signing / recompute on trusted servers) — different method, same trust problem.

Editor takeaways (actionable)

Google K-12: decide before early 2027 whether your district accepts Gemini-class Search, switches discovery vendors, or drafts board language for a Search outage — “AI off” is no longer free.
ChatGPT vs Google search: never cite the ~20% figure without Fry’s web-search-trigger definition; track referrals separately from in-chat search events.
FlashProve: watch for API attestations and auditor tooling; do not assume Claude traffic is zk-proved today.
wordok angles: education policy posts should cover Search-as-AI bundling; SEO posts should teach metric hygiene; security posts can treat provable inference as the next trust checkbox after SSO and data-residency.

Primary sources

The New York Times / AP / The Verge / Ars Technica — reporting on Google K-12 Gemini & Search default-off (Nov 3, 2026); see e.g. The Verge and Ars Technica
Big Technology — OpenAI: ChatGPT search ~20% of Google’s; WAU growth (Adam Fry interview)
The Verge — OpenAI says ChatGPT is catching up to Google in search
Anthropic — Progress on provable inference (FlashProve) (Nov 2, 2026)
GitHub — anthropics/FlashProve

Topic selection may overlap other AI digests on the same calendar day. Wording, sectioning, and recommendations here are written for wordok.top. Prior issue: 2026-11-03.