Google Bundles AI and Search Into Schools: A Battle Over Educational Control

Google recently made a decision that has drawn attention from both the education and tech worlds: in school settings, AI and search are packaged under a single switch. Schools must either accept both Gemini and search summaries, or lose access to search itself. This seemingly technical product decision actually surfaces a deeper question in the AI era—when tools start thinking for us, who decides how children should use them?

Bundling Strategy: No Separate Switch

According to an announcement by Chris Phillips, Google's Vice President of Education, on the corporate blog, Gemini study notebooks, Gemini in Classroom, and the search engine itself will be off by default in K-12 student accounts until school administrators actively enable them. Details will reach schools over the coming months, with deployment to follow.

On the surface this looks like a concession—after all, Google had previously enabled Gemini by default in Classroom, which serves over 150 million students and teachers. But the key detail Phillips shared with The New York Times is this: there will be no AI switch independent of search, because the two are "deeply integrated." That means schools that decline AI summaries will have their students see a notice that search is unavailable.

This logic has provoked strong backlash. Senator Bernie Sanders called it "extortion," and the New York City mayor stated that "no company should decide how children are taught." Google's response: this gives schools more choices, not fewer.

The Technical Loophole: Separation Already Exists

Google's claim of "deep integration" runs into an unavoidable technical fact. According to the program's analysis, since May 2024 Google Search itself has included a filter called "web" that shows only links and no AI summaries. In other words, the separation of search and summaries has already been achieved technically—what's missing is not capability, but the willingness to offer that option to schools.

It's like a supermarket hiding loose fruit on a warehouse shelf and telling schools they can only buy it bundled with pastries. Google emphasizes the indivisibility of integration while retaining a pure-link mode in its consumer product—this asymmetry itself speaks to the weight of commercial considerations.

Google Search's "web" filter is a feature launched in May 2024. Users can switch to the "Web" view in the tab bar at the top of the search results page, which returns only traditional blue links without AI Overviews summaries, knowledge panels, or featured snippets and other generative content. This filter was initially made available to all users globally and was seen as a response to concerns about declining publisher traffic. Architecturally, AI summaries and the search index are decoupled: the summary layer runs on top of retrieval results and can be switched on or off independently. So "deep integration" does not mean indivisibility at the code level, but rather a binding at the level of product experience and business strategy. If schools request that summaries be disabled, technically only the summary layer needs to be turned off, without touching the search index. This fact weakens the technical basis of Google's bundling argument.

Evidence for Learning Mode: What the Data Says and What It Leaves Out

Google paired this bundling with an educational argument: its "Guided Learning" mode does not give answers directly but assists thinking through questions and step-by-step breakdowns. The feature has rolled out globally in Classroom and will be tested in search-side student accounts in some U.S. schools.

The supporting evidence comes from one of Google's own experiments: over 300 University of Tokyo students used the Gemini app (not the search engine) to study a statistics course. During a half-day of learning, half of them additionally used guided tutoring. Results showed that students using only conventional resources improved by about 5 points on the post-test, while those using tutoring improved by about 8 points.

The gap is not large, and Google did not disclose starting scores or state whether the difference is statistically significant. More importantly, it creates tension with another study: as the program previously reported, Patrick O'Neill's preliminary meta-analysis of 44 experiments showed that AI's average effect on learning outcomes is close to zero.

Both can be true at once. AI that gives answers directly removes the effort required for learning, while AI that forces thinking may help slightly. But even guided mode in search only triggers when Google "detects this is homework," whereas direct answers are always just one click away. Which will a tired student at 11 p.m. choose when both are right in front of them?

Guided Learning is an interaction mode introduced by Google Gemini that does not give the final answer directly but guides users to derive it themselves through follow-up questions, hints, and step-by-step breakdowns. Its design philosophy draws on Socratic questioning and the "generation effect" in cognitive science—learners remember better when they actively produce answers. But triggering this mode on the search side depends on Google's automatic identification of "homework queries," which means it is not the default path; users must actively choose it or the system must determine it is a learning scenario. Direct answers are available in a single click, while guided mode requires multiple rounds of interaction—the asymmetry in user effort is the backdrop for understanding the experimental results.

Data and Contracts: The Ignored Core Issue

For school administrators, this decision is no longer a technical matter. They must choose between "a tool students use to skip practice" and "losing the world's most-used search engine." But the announcement completely omits commitments on student data use—which is precisely the clause schools should be fighting for first in contracts.

If schools demand both an independent switch and data terms, the form that is technically unavailable today often comes into being. For families, the lesson from the experiment is practical: the key is not whether to use AI, but whether the tool is asking questions or answering them. This can be observed at home.

Google's move reveals its value judgment: it does not want a generation to learn how to search without summaries, because today's students are tomorrow's search users.

AI2's StateBench: Why Long-Context State Tracking Defeats Top Models

The program also reported on StateBench, released by the Allen Institute for AI (AI2), a benchmark measuring a model's ability to track state changes in long texts. Its design reveals some truths about how models work internally.

The tests have four difficulty tiers: the simplest is when a value is set once and never changes (such as a birth date); next is when a value is replaced and only the latest value matters (such as a thermostat setting); the third level is when each change accumulates on the previous value (such as a basketball score), with no step skippable; and the highest level is when variables are assigned to each other, like a shell game where a ball moves among cups and every swap must be tracked to determine the final position.

This classification is not arbitrary but derives from the Crony-Rhodes theory in algebra, which decomposes such problems into basic components.

The results largely match the theory: Gemini, Claude, and OpenAI models perform well on the first two difficulty levels but decline sharply at high difficulty and long context, with accuracy at the highest difficulty dropping to very low levels. Architecture alone is not enough—AI2's Olmo hybrid model, despite containing recurrent layers, performs no better than random guessing at the highest difficulty. Only when AI2 specifically trained a small model did a more flexible recurrent variant solve problems far beyond the training distribution length, which Transformer never learned.

This means recurrent architectures provide the "notebook," but someone must teach the model how to use it—conventional training cannot do this. Format also changes results: the same problem presented as spreadsheet editing causes a small model that could otherwise solve it to fail.

The implications for practitioners are concrete: contract amendments, tables where cells reference each other, inventory where items move between warehouses—these look like reading problems but are actually high-difficulty state tracking. In such cases, require the model to write out states step by step, or use code, rather than relying on one-pass reading. The longer the document, the more vigilance is needed.

Crony-Rhodes theory is a classic result in algebraic automata theory, proposed by Larry Crony and John Rhodes in the 1960s. It decomposes finite-state automata into three basic components: reset automata (forcing a state to a value), permutation automata (one-to-one mapping between states), and unit automata (state remains unchanged). StateBench's four difficulty tiers correspond precisely to the combinatorial complexity of these components: single setting corresponds to reset, replacement to simple permutation, accumulation to a counter requiring memory, and mutual assignment to a complex permutation group. The significance of the theory is that it provides an "atomic" classification of state-tracking problems, giving the benchmark's difficulty divisions a mathematical basis rather than an empirical guess. This also explains why Transformer collapses at the highest difficulty: its feedforward architecture lacks recurrent state memory and struggles to simulate the composite operations of permutation groups.

FlashProof: Letting Anyone Verify Model Identity

Anthropic, in collaboration with cryptographers from USANA Institute of Technology and MIT, released FlashProof, an open-source system that uses succinct zero-knowledge proofs to verify that a given model actually generated a given response.

The principle rests on an old idea: verifying a computation can be much cheaper than performing it, just as division can be checked by multiplication. The provider publishes a sealed fingerprint of the weights and attaches a certificate to each response proving that every word is what the model would choose. The client verifies on their own computer without ever seeing the weights.

This is feasible because the Llama model used in testing is a mixture-of-experts architecture—only a small fraction of total parameters participate in computation each time, so only that portion needs to be proven correct. A single H100 generates proofs at about 100 tokens per second, and verification takes about 17 seconds on an ordinary processor.

The limitations are equally clear: generating proofs costs dozens of times more than directly serving responses, so the positioning is post-hoc spot-checking, like a cashier checking some receipts. It currently applies only to contexts of about 1,000 words, far below real conversation lengths, and supports only model versions that always select the most likely token, excluding the random sampling used in real services.

The authors suggest regulators could require certificates for served content and verify that only approved models are running. But they acknowledge this cannot prove those GPUs aren't simultaneously training something else. The protocol has been formally verified in Lean, but the fast verifier used in daily practice does not carry the same guarantees. This is a research prototype, which Anthropic lists as one of several paths to ensuring service integrity.

A succinct zero-knowledge proof is a cryptographic protocol that allows a prover to convince a verifier that a computational statement is true without revealing any information beyond "the statement is true," with proof size and verification time far smaller than re-executing the computation. FlashProof applies this to model inference verification: the prover must prove that "given the weight fingerprint and input, the output token sequence is the result of the model's greedy decoding." The mixture-of-experts (MoE) architecture is key here—each forward pass activates only a small subset of experts among total parameters, so the proof only needs to cover the activated subnetwork, greatly reducing proof cost. But MoE routing decisions themselves also need to be proven, which increases protocol complexity. The scheme currently supports only greedy decoding, because the probabilistic path of random sampling is difficult to express in a succinct circuit.

Palantir's Transformation Signal: Enterprise Customers Are Changing AI Procurement

Palantir's quarterly report presented a different picture for the first time: it looks more like an enterprise software company than a defense contractor. According to its SEC filing, U.S. enterprise customer sales grew significantly year-over-year and slightly exceeded U.S. government revenue.

What they sell is the ability to use frontier models or train open-source models without moving data out of internal systems, along with on-site engineers. The key is contract structure: total new U.S. enterprise contract value was considerable but noticeably lower than the previous quarter, even as it grew year-over-year. CEO Alex Karp acknowledged that customers now trial faster and sign shorter contracts, defending this as an entry point for later expansion.

This is precisely the prevailing pattern for enterprise AI this year: bypass bureaucratic processes to start quickly, and leave just as quickly. For a company valued on multi-year revenue, the difference between a customer staying ten years versus trialing for six months is fundamental. Palantir itself is betting on the former and has raised its full-year U.S. enterprise sales forecast.

Data Center Power Crunch: Retired Aircraft Engines Take the Stage

According to The New York Times, grid connection wait times have grown so long that some data centers are turning to retired aircraft engines to generate their own power. ProEnergy converts retired engines used by GE and others in Boeing aircraft into gas generator sets. According to International Energy Agency data, connecting 100 megawatts of power in Northern Virginia takes about seven years, while a data center using these engines can be operational in about two years.

The cost is physical: aircraft engines do not recover exhaust waste heat (unlike combined-cycle plants), so they burn more natural gas and emit more per megawatt. SpaceX already operates Musk's Colossus data center and has raised its 2027 projection to about 13 gigawatts, equivalent to 13 nuclear reactors. One analysis firm identified nine projects under construction or in operation and noted that power supply sometimes appears only after construction begins.

There is also a detail linking two news stories: under a commitment signed in July, the generators powering that center must be reduced to less than half power by 2027. For neighboring communities, this means engines that never appeared on planning maps will migrate as permits expire.

Open-Source Models Can "Read" Weights: The Experimental Barrier Is Lowering

Earlier work from MIT showed that frontier models can infer the training objective of small neural networks by reading their raw weights. Simon Willison reproduced the experiment on his own laptop using the open-source Qwen model (27 billion parameters, about 17GB): given raw numbers from four networks trained for simplified classic games, without a calculator or code, it guessed three of four correctly—Pac-Man, Pong, and Breakout—mistaking the obscure game Freeway for Pac-Man. Each game took 45 minutes to over an hour. Claude, through its service, solved Breakout in under 10 minutes.

Skepticism exists as well: some developers criticized the work for conflating "reading weights without tools" with "modifying weights with code," and others argued the so-called breakthrough is just a step any researcher would take. But for researchers with limited budgets, the signal is clear: an experiment that yesterday required a frontier model can today be reproduced slowly and for free at home.

Nexus Mods: Mandatory AI Labeling, but Honesty Has a Cost

Nexus Mods, the largest PC game mod repository, already had AI content tags; the new rule makes them mandatory. Uploaders must declare whether they used AI, and if used as an assistive tool, must disclose where and how. Existing uploads must comply by January 2.

Site owner Victor Folman refused to ban AI, for a noteworthy reason: AI has permeated the entire process, from code autocompletion to textures and voice, and a ban would not eliminate it—it would only mean no one admits to it. He also pledged not to train models on mods or sell them to those who would. This is a "ingredients list" model rather than a "customs" model—not deciding for users, but providing information.

But the same site introduced a filter to hide all labeled content, starting with game pages and the homepage, then extending to search. So declaring may cost downloads from day one. The honesty it demands has a price, and those who pay it are the ones practicing honesty.

A search engine that won't sell search separately, a certificate anyone can verify but no one can see the weights behind, a label that requires every author to self-declare—"verification" is shifting from a technical problem into a form of power.

One-line Summary

Google's bundling of AI and search sparks controversy, while multiple AI advances reveal structural limits of verification and tracking.

Paragraph Summary

Google bundles Gemini and search into a single switch in schools, preventing schools from disabling AI summaries independently—even though search already has a links-only filter mode—a move critics call "extortion." Google defends the bundle with a guided-learning experiment, but gains are modest and not statistically significant, creating tension with meta-analyses showing near-zero AI learning effects. Meanwhile, AI2's StateBench reveals structural deficiencies in Transformer long-chain state tracking, which recurrent architectures can only弥补 with specialized training; Anthropic's FlashProof uses zero-knowledge proofs for model identity verification but is costly and limited to short contexts. Palantir's quarterly report shows enterprise AI procurement shifting to fast trials and short contracts, data center power shortages force retired aircraft engines into service, and MIT experiments show open-source models can read weights to infer training objectives. Nexus Mods mandates AI labeling but honesty brings download losses. The core issue: verification is evolving from a technical problem into a form of power.