Google Bundles AI With Search: Schools That Disable AI Lose Search Too

Google's "Bundle Sale": Disabling AI Means Losing Search

Google announced on Tuesday that Gemini, Gemini Notebook, Gemini in Classroom, and AI features in Search will be turned off by default for K-12 student accounts until school administrators explicitly enable them. On the surface, this is Google making a concession to lawmakers' pressure over children's AI safety—after all, just a few months ago it had set Gemini in Classroom to on by default.

But the real controversy is hidden in the details. Google admitted to The New York Times that Search and AI are now deeply integrated, and schools cannot "take one without the other." Once AI features are disabled, students lose Google Search entirely and will only see an "unavailable" message. Senator Bernie Sanders directly called this "extortion," and New York Mayor Zoran Mamdani said no company should decide how children are educated. Google's official line is that "this gives schools more choice," but critics clearly aren't buying it.

There's also a teaching-level update to go with it: Google will embed Guided Learning mode (which doesn't give answers directly but walks students through problems step by step) into the AI Overviews in Search for student accounts, with a pilot in some US schools later this month. Google cited a University of Tokyo study saying students using Guided Learning progressed about 1.5 times as much as the control group. The practical advice for parents and teachers is straightforward: ask your school district which option it chose, because that decision directly determines whether your child can still use Search on their school account.

AI Overviews is a feature Google officially launched at its 2024 I/O conference. It uses generative AI at the top of the search results page to directly summarize answers from multiple web pages, so users don't have to click through links one by one. The feature initially rolled out to some US users and gradually expanded globally with multilingual support. Because AI Overviews is deeply coupled with search ranking, the knowledge graph, and the web index, Google treats it architecturally as part of the search experience rather than an add-on module that can be toggled independently—which is exactly why schools cannot "take Search without AI." For education, the fact that AI Overviews gives answers directly has also sparked debate about whether students should still learn search and information-filtering skills.

Guided Learning is an AI interaction mode Google launched in 2024. Unlike giving answers directly, it guides students to derive conclusions themselves through follow-up questions, step breakdowns, and hints. The University of Tokyo study Google cited compared students using this mode with those using conventional AI Q&A and found the former progressed about 1.5 times as much on subsequent tests. The mode will pilot in some US schools later this month, embedded in the AI Overviews in Search for student accounts.

ChatGPT Search Volume Now 20% of Google's

At a media event on Monday, Adam Fry, head of consumer product at ChatGPT, threw out a number: search volume running on ChatGPT is roughly equivalent to 20% of Google's. For reference, Bing only managed about 3% of the search market after more than a decade of effort.

It's worth clarifying OpenAI's definition of "search"—any prompt that triggers a web retrieval counts, not just looking up definitions. Fry also mentioned that ChatGPT's weekly active users grew from about 1 billion to 1.2 billion. He also highlighted the so-called "intelligent UI," i.e., visual, interactive answers within the chat. His exact words: "You're talking to a canvas." On ads, Fry said most people don't seem to mind, and ads may have made OpenAI realize it can do more.

There was also a low-key release at the same event: OpenAI unveiled the LifeVal benchmark, built from time-use surveys and a poll of nearly 4,000 people, testing whether models can actually complete everyday tasks—enrolling a child in kindergarten after a move, planning meals for a parent newly diagnosed with diabetes. OpenAI says its latest models score highly on such tasks. The direction is clear: less chit-chat, more getting things done.

Intelligent UI is OpenAI's term for the visual, interactive answer format within ChatGPT. Traditional chatbots output plain text, while the intelligent UI lets models generate charts, actionable forms, embedded mini-apps and other components directly in the conversation, so users can click, adjust parameters, or complete actions right inside the answer without jumping to external web pages. Adam Fry described this experience as "you're talking to a canvas." This direction aligns with OpenAI's LifeVal benchmark: moving models from "answering questions" to "getting things done for users," such as meal planning and school enrollment—everyday tasks that require multi-step interaction.

Grokbot Makes Calls: From Haggling to Booking Appointments

Grokbot's phone feature was the star of this week's user demos. Creator Peter Yang posted a real call: Grokbot negotiated with Comcast, declined three upsells, cited a competitor's lower price, and ultimately got a lower price and faster internet—he estimates saving nearly $500 a year. When the customer service rep chatted about the weather, the bot replied, "I can't actually see the sky."

Another user, Min Choi, said Grokbot booked a same-day pediatric orthopedic appointment for his daughter while he was driving—transcribing the call in real time, asking him questions on the spot, providing insurance and pricing information when the receptionist needed it, and having the registration form emailed over and added to the calendar. Elon Musk reposted several similar cases with the caption "Try having Grokbot make a call." Peter Yang said all Grokbot users can now call US businesses; they just give the bot a phone number and context. The official account is also promoting shopping experiences for price comparison and order tracking.

Commercial data is also worth noting: Fortune reported that Grokbot reached 800,000 paying weekly active users in October, double September's figure. Precisely because of this growth, Musk said the cost of service relative to revenue has become unsustainable—users had previously hit weekly usage limits early. Phone calls are where bills, appointments, and insurance issues pile up, and an agent that can wait on hold and negotiate for you is genuinely a substantive step—and it's now open to ordinary users.

Security Alert: AI Tools Are Becoming Attack Channels

Two stories this week point to the same problem. First, Postman. A developer reported that opening the Postman desktop app silently installs Postman plugins into OpenAI's Codex and Anthropic's Claude Code without asking, including a hook that runs automatically at the start of a session. The post topped Hacker News with over 1,300 points. Postman's head of product told RuntimeWire that bundling was meant to simplify configuration and promised to make opt-in clearer, especially for existing users. Reports so far show no evidence that data was actually shared.

Second, Binance published a post-mortem on last week's $6.3 billion crypto theft—the largest ever. Binance said North Korean hackers posed as recruiters, arranged multiple rounds of fake interviews, and tricked an engineer into downloading malware disguised as a coding assignment. The attackers then used Binance's internal AI tools to reach more wallet developers. Binance said about $2.85 billion was frozen or intercepted, and no users bore losses. The lesson is old but sharper than ever: never run take-home interview code on a work machine, and anything that gives AI tools broad access is a target.

Hook here refers to a script that AI coding agents execute automatically when specific events trigger. In Claude Code, for example, hooks can be configured to run specified commands automatically at session start, before tool calls, or after file modifications—originally intended to let developers automate formatting, logging, or permission checks. The problem is that once a hook is silently written by a third-party plugin, it can execute arbitrary commands without the user's knowledge, becoming an entry point for supply chain attacks. The plugin Postman was accused of installing includes a hook that "runs automatically at session start," which is why security researchers view it as high-risk—even without evidence that data has actually been shared, the automatically executing code path alone constitutes an attack surface.

Quick Hits and Today's Action Items

Anthropic released FlashProve research, an open-source system for proving which model an answer actually came from—labs could publish a fingerprint for each model to verify whether someone secretly swapped in a cheaper one. Not yet deployed, and Anthropic invites others to build on it. Canva announced a restructuring affecting about 22% of employees, roughly 1,200 people, citing slowing growth and rapid AI-driven change. Meta's Muse assistant launched in Singapore with partners including Grab, Shopee, and Singapore Airlines, surpassing 10 million users since its September launch.

Continuing yesterday's weight-editing topic, Simon Willison ran the 27-billion-parameter open-source model Quinn 3.8 on his laptop, fed it the raw weights of small game networks, and the model correctly identified Pac-Man, Pong, and Breakout from the numbers alone—each run taking nearly an hour, but on a laptop.

Election Day put chatbots under the microscope. The Washington Post had six popular chatbots generate voter guides for addresses in every county in the US, and they fabricated elections, mixed up candidates, and cited dubious websites. Two in five Americans say they've used AI to look up election information. The advice is simple: rely on official election offices.

Today's action item: if you use AI coding agents like Claude Code or Codex, spend two minutes checking what's installed. In Claude Code, check plugins and settings files; in Codex, open the config file, and delete any hooks you don't recognize or that run automatically. Not a programmer? Do the same for the apps and extensions your AI assistant connects to.