Beyond an AI Ban or Bundle: The Learner Agent

In September, New York and Los Angeles moved to restrict AI systems on school devices. Google’s announcement earlier this week can be seen as a response. But it also told America’s schools, in effect: take our AI, or lose our Search. Then it spent a day insisting that is not quite what it said. Both the ultimatum and the walkback are worth a closer look. Together they reveal something the current fight over AI in classrooms keeps overlooking.

On Tuesday, Google announced that for K-12 student accounts on its education software, Gemini, Gemini Notebook, Gemini in Classroom, and Google Search would all be switched off by default until a school administrator turns them on (NBC News). Asked by the New York Times why Search was in the bundle, Google’s education chief Chris Phillips explained that “Search and AI are deeply integrated,” so students at schools that decline AI “will not have access to search” (TechCrunch).

The reaction was fast and unusually broad. Senator Bernie Sanders called it extortion. New York Mayor Zohran Mamdani accused the company of using monopoly power to push AI into public schools (CNBC). Fairplay’s Josh Golin said the attempt to strong-arm schools would backfire, and urged every district to switch Google’s AI off the moment it could (AP via ABC News).

By that evening Google had clarified: schools may turn Search back on by itself, without Gemini. But note what “Search by itself” now means. Search comes with AI Overviews, and the same announcement unveiled a new AI feature inside Search to guide student learning (Engadget). The explicit ultimatum is gone. The integration that produced it is not.

The wrong axis

Everyone in this fight, Google included, is arguing along a single axis: AI on, or AI off. Google wants it on, by default and everywhere. Its critics want it off, and the most forceful of them want it off for good. Fairplay’s position is the logical endpoint: AI is not safe for kids, so turn it off the minute you can.

That is an understandable reaction. It is also the wrong axis. The binary on/off debate quietly accepts the premise Google most wants to have accepted: that the only AI a student will ever encounter is the vendor’s AI. One built to the vendor’s purposes, with the vendor’s business model sitting behind the screen. Once you accept that premise, the only question left is how much of it to allow. Abstinence becomes the sole alternative to capture. It is a binary that Google is betting on winning.

The better question is not whether, but whose. Whose interests does the AI in the classroom actually serve? Who does it answer to, when the learner’s interest and the vendor’s interest inevitably diverge, as when a company’s product metrics meet a twelve-year-old’s attention? An AI that is on the learner’s side is a different thing from an AI that is merely switched on, and a different thing from an AI that is switched off. That third option is the one nobody in this week’s news cycle has put on the table.

A kill switch is not agency

Google’s defense has been the word “control.” Schools told us they want control, the company said, so we are giving them more choice, not less. And it is true that an administrator can now flip Gemini off.

But look at what kind of control that is. It is a switch over Google’s product, with the default set by Google, the integration decided by Google, and the terms written by Google. The administrator gets to choose between two of Google’s configurations. The student, the parent, and the teacher do not get to choose anything. They are the objects of the decision, not parties to it.

Integration is the lock-in

“Search and AI are deeply integrated.” Mr. Phillips offered that as a technical fact, the way one might say the engine and the transmission are deeply integrated. But that is a design choice, and a very recent one. Search worked without AI Overviews until May 2024. The decision to fuse them, and then to make the fused product the only one on offer to schools, was made down the street from where I live, in Mountain View.

That choice is the real story, and it will outlast this week’s walkback. Once the AI is woven into the infrastructure a school cannot do without, the school’s “choice” is no longer about AI at all. It is about whether to keep the search engine, the email, the documents, the classroom software, and the laptops. For millions of students who are required to use Chromebooks and Google’s education tools, switching search engines is not a simple matter (TechCrunch). That dependency is the lever, and the lever is what the senator and the mayor were really objecting to, whether or not they used the word.

In my 2024 book Reweaving the Web, I wrote about the layers of dispossession that platforms build, one on top of another, until a person’s data, attention, and choices all belong to someone else. What we saw this week is that pattern arriving in the schoolhouse as well, with the student’s learning itself as the thing being integrated. The lesson is not that Google (my former employer) is uniquely villainous. It is that AI delivered from inside the platform stack will always answer to that stack. If we want an AI that answers to the learner, it has to live somewhere else.

The third path: an AI on the learner’s side

There is a pivot I have called “ANDing.” Constrain the tech platforms, yes. But also empower the rest of us. We can do both. In the age of AI, ANDing is no longer optional.

At GliaNet we call that human-side counterpart a Personal Augmentation Agent. In the classroom, we call it the Learner Agent. The idea is simple to state. The learner, not the vendor, gets an AI. It belongs to the student, works for the student, and is bound by enforceable duties to the student (for minors, under parental direction). It can talk to Google’s tools, or anyone else’s, but it does so as the learner’s representative, not as the platform’s front door. And it carries the learner’s portable memory files from one learning situation to the next, private and protected.

Three duties make it different in kind from what Google is offering:

• A duty of care. The agent must be competent and must not harm the learner. For a student, that means it is built to help them learn, rather than to keep them engaged. It asks before it answers. It never does the assignment; it helps the student do it. This can include offering what we have called “loyal friction”: the productive resistance a good tutor supplies, rather than the frictionless answer a product optimized for usage supplies. The emerging evidence on “tutor mode” AI, which guides rather than completes, points the same way. A duty of care is what keeps a tutor in tutor mode when the platform metrics argue otherwise.
• A duty of fidelity. The agent keeps the learner’s confidences. What a child asks, struggles with, or gets wrong is not raw material for a profile, a product roadmap, or an advertiser. It is the learner’s own, held the way a school counselor or a librarian holds it.
• A duty of loyalty. When the learner’s interest and anyone else’s interest collide, the agent sides with the learner. Not the district, not the vendor, not the agent’s own maker. That is the one duty no platform AI can offer, because a platform AI’s loyalty is already spoken for.

These duties are not promises buried in a terms-of-service page. In the GliaNet model they are embedded in the agent’s architecture and backed by a real entity, a Net Fiduciary™, that is accountable when the agent gets it wrong. Someone is on the hook. Recourse exists. That is the difference between an AI that is “safe for kids” as a marketing claim and one that is safe for kids as a structural, accountable fact.

Working for the student, not the school or the platform

Consider who is represented when AI enters a classroom. The vendor is there to protect its business model. The district is there to manage its budget, its obligations, and its liability. The only party without anyone acting for it is the student, the one whose learning is actually at stake. Children are the people least able to recognize when a system is working against them, and least able to push back when it is. The Learner Agent exists to give the student someone acting for them.

It starts with knowing the learner. Over time, the agent comes to understand how a particular student learns: where they get stuck, what kinds of explanations work for them, when they need a hint, and when they need to struggle a little longer. A platform uses that kind of knowledge to keep a student engaged. The Learner Agent uses it to help the student understand the material. It slows down when a concept has not landed, keeps the work with the student instead of doing it for them, and measures success by what the student can now do on their own, not by how long they stayed on the screen. That record belongs to the student and moves with them, grade to grade and school to school, under the family’s control.

None of this sets the Learner Agent against the school. Schools owe their students a proper education, and the agent supports that work. It follows the assignment, honors the academic integrity rules, and works with whatever software the classroom uses. But a school’s job is to teach every student in the building. It cannot also stand beside each child as that child’s own representative. The Learner Agent fills that role, as an advocate for one learner, working alongside the institution responsible for all of them.

Where does such an agent live? Not necessarily in the school district’s procurement office. One natural home is the public library. Libraries are trusted civic institutions with a long tradition of patron confidentiality, serving every learner in town, not only the ones enrolled in a given school. A Learner Agent with every library card is a far more fitting arrangement than an agent bundled with every Chromebook. School-based pilots belong in the picture as well, especially where states are already building learner-controlled records. In either setting the point is the same: the agent travels with the learner, from home to school to library, and answers to the learner wherever it goes.

Taking the safety objection seriously

The people saying “turn it all off” deserve a real answer, not a slogan. Their worry is well founded. Much of the AI being pushed into schools is built to maximize use, collects more than it needs, and makes it easy for students to skip the thinking that learning requires. If the only AI available is that version, turning it off would be a defensible call.

The Learner Agent’s answer is not “trust us.” It is that the objection describes platform AI, and platform AI is precisely what a fiduciary agent is designed not to be. No engagement optimization, because the duty of care forbids it. No data harvesting, because the duty of fidelity forbids it. No quietly siding with the vendor, because the duty of loyalty forbids it. And an accountable entity behind the agent that can be held to those duties, through its own binding commitments and the student-privacy rules schools already live under.

How these Learner Agents are designed is also a fair question. The honest answer is that no technology company or nonprofit, GliaNet included, knows enough about how children learn to design such agents alone. So our core duties of care, fidelity, and loyalty would be joined to detailed design principles written by the people who do know: parents, educators, librarians, researchers, and other domain experts. Their aim would be to shape, openly and transparently, the right tutor agent for specific classroom situations and student needs. A parent could see what the agent remembers about their child, and correct or erase it. Parents also would be free to opt in, or not, to what is produced. And there would never be an advertiser in the room.

A critic is entitled to say: show me. That is fair, and it is why GliaNet is pursuing pilots rather than proclamations. But the right response to a bundled, vendor-loyal AI is not to leave the field. It is to put a learner-loyal AI on it.

And of course: who pays?

Start by noticing who pays now. Google’s classroom AI is free the way a loyalty card is free: the student pays, just not at the checkout. The price is paid in attention captured, data collected, and learning shaped toward a product’s metrics, rather than a child’s understanding. That is a subsidy, drawn against the student’s own well-being, and once you see it you can’t keep calling the alternative “too expensive.” The visible cost of a Learner Agent is modest. The computing could run to as little as tens of dollars per learner per year. The portable memory feature adds a bit more. The all-in cost, with a human being accountable behind the agent, is in the range districts already pay for commercial AI tutors. What costs money is the duty, the accountable institution standing behind the agent -- and libraries and school systems already are that institution.

So the funding can be ordinary where ordinary works: public budgets, existing edtech dollars redirected by a procurement rule, philanthropy for the first mile, a modest family tier above a free public one. And where ordinary falls short, we should be willing to get creative, because the harm we are replacing is not ordinary either. A per-student contribution from the platforms whose integrations are creating the problem. A share of any settlements from related lawsuits. Library consortia, parent cooperatives, state AI-in-education funds. One rule governs all of it: the agent’s knowledge of the child is never for sale, to anyone, in any form. And one principle: the payer is not the principal. A Learner Agent works for the learner, whoever writes the check, and that is written into its charter, rather than left to anyone’s goodwill.

What to ask for now

This week’s public backlash creates an opening. In a single week, a senator, a big-city mayor, and a leading children’s advocacy group have all said that the AI being offered to schools serves the company that makes it. The next step is to say what should be offered instead.

• Parents can ask their district a single question: whose side is this AI on? If the answer is a settings page, that is not an answer. Ask whether any option exists that is bound to serve your child, and whether the district has looked.
• Teachers can insist on tutor-mode behavior as a condition of any AI in their classroom: an assistant that asks before it answers, and leaves the learning with the learner. And they can ask who is accountable when it fails.
• Districts can write fiduciary duties into procurement. Not “AI off,” but “AI that owes us a duty of care, fidelity, and loyalty, with a named entity behind it.” Vendors respond to RFPs. Make the duties a line item.
• Libraries can step forward. The public library is the community-based institution best placed to offer a Learner Agent to every resident, child or adult, outside the platform stack entirely. Library systems interested in a pilot should get in touch with us.
• Lawmakers can recognize that the fix for coerced bundling is not only antitrust after the fact. It is a legal floor of duties for any AI that works with children, and a protected lane for agents that work for them.

Giving students a deeper edge

Google will keep saying it listened. In a narrow sense it did: the switch now exists. But the choice it offers is still a choice among its own pre-defined configurations, and the integration that made Search and AI inseparable is still the default shape of the product. That is the deeper edge of this story, which will still be here after the headlines have moved on.

The students sitting in front of those Chromebooks are not going to grow up in a world without AI. The question is whether they grow up with an AI that works on them, or one that works for them. Banning the first does not produce the second. Building the second is the work. Not ban, and not bundle. Give our students an agent of their own, on their side, with duties that hold.

Want to take this further?

If you are a parent, educator, librarian, or district leader who wants to see what a Learner Agent could look like in practice, we would like to hear from you.

Contact Us