Common Sense Media’s verdict is blunt: Google’s AI Mode integrated into Search poses an “unacceptable risk” to young users. The recommendation is drastic — stop using it altogether, at least until schools can disable the AI features. But there’s a catch that turns a typical education alert into a structural problem: Google can’t turn it off. Not for a single school, not for a district, not on request.

The report documents three critical behaviors. The AI completes homework assignments for students, erasing the learning effort. It repeats false information with a confident tone that makes it believable, bypassing any educational verification. And it exposes minors to inappropriate content, a risk the watchdog believes cannot be contained with current controls. The conclusion: as long as the AI stays enabled, every student connected to Google Search operates in a grey zone of responsibility.

Why can’t Google disable it? AI Mode is stitched into the search engine’s architecture, running on cloud servers that handle billions of daily queries. There’s no kill switch for institutional users — personalization goes through managed accounts, but the deep AI features are active at the backend level and don’t allow granular exclusions. It’s the epitome of a cloud model where the service provider decides what’s available, and the consumer — even an entire school infrastructure — can only accept or reject the whole package.

Here lies the short circuit. A school can’t just walk away from Google Search, which has become a de facto utility. The contradiction exposes a broader sovereignty issue: cloud-based tools shift the locus of control far from the end user, turning security policies reactive rather than preventive. In education, where protecting minors is also a legal duty, offloading everything to an external provider means accepting a structural misalignment between one’s responsibilities and the means to enforce them.

The flip side concerns those working on local stacks. This story illustrates why environments where the LLM runs on organization-controlled hardware allow for access management, filtering, and shutdown with a granularity impossible in the cloud. It’s not an isolated example: organizations with strict compliance needs are already exploring self-hosted inference architectures, where the ability to remove a feature or monitor output tokens doesn’t depend on an external vendor’s roadmap. The trade-offs are familiar — operational complexity, TCO, internal expertise requirements — but AI-RADAR offers analytical frameworks (see /llm-onpremise) to weigh those balances without rushing to prescriptive advice.

The conflict reaches beyond Google and homework. On one side, big tech accelerates the release of AI features into mass-market products to stay competitive; on the other, the institutions using them in sensitive contexts are left without off switches. The paradox is that the more AI becomes pervasive in areas like education, the more urgent it is to provide local control filters. This isn’t just about silencing a chatbot — it’s about returning the power to decide what enters a classroom to those who run that classroom. While regulation lags, the Common Sense Media case is a litmus test: if a giant can’t guarantee deactivation today, tomorrow the clash between centralized models and sovereignty demands will only intensify.