The Feed Is Already on Your Mind

The feed knew what to put in front of me before I remembered why I had opened it.

I had entered with a question. Nothing dramatic. Something I wanted to understand well enough to use. A few minutes later, the question was gone and I was moving through answers to desires I had not declared: novelty, recognition, irritation, relief, the small animal hope that the next swipe might contain the thing that made the previous swipes feel necessary.

This is usually described as distraction. That is a lazy and passive description. The feed had not failed to serve my intention. It had replaced the intention with a better-predicted appetite, with an even better prediction about what I craved.

I was not standing outside the system making a clean accusation against weak people and manipulative companies. I was in it, wet to the skin, participating in the bargain. The feed was learning me faster than I was learning anything inside it. It had already won but I failed to realize it that day. It terrified me when I actually realized it.

We won the old fight for access to information. We can now reach more writing, instruction, demonstration, argument, research, and recorded experience than any person could absorb. The victory is real. So is the new problem it created.

Information became abundant before we learned how not to drown in it.

Access was the old fight

For most of history, access was a defensible moral center. Knowledge was scarce, expensive, geographically trapped, institutionally guarded, or written in a language the learner was not allowed to enter. Libraries, schools, public education, publishing, search, and the open web each widened the well.

That work should not be diminished merely because abundance later developed its own hazards. A person with a connection can now learn from a lecture on another continent, inspect a paper the day it appears, hear the dissenter an institution ignored, or acquire a practical skill without first being admitted to the room where it was traditionally taught.

But a resource changes character when its supply exceeds the ordinary capacity to judge and absorb it.

Information is water: necessary to live, dangerous when unfiltered, destructive when flooding. Scarcity makes access the urgent engineering problem. Abundance makes filtration, pressure, channeling, storage, contamination, and flood control equally urgent.

The internet kept solving for the well after the flood had arrived.

Search could find almost anything. Distribution could deliver it instantly. Storage could preserve more than a person would revisit in a lifetime. What remained unsolved was the learner’s continuity: why they began, which sources deserved trust, what order would build understanding, what should be ignored, what had been misunderstood, and where to return after appetite carried them away.

Stewardship is not a polite synonym for restriction. It is what allows a resource to remain useful after abundance changes its risks.

The curriculum that calls itself entertainment

The feed already performs stewardship. It simply refuses the obligation while keeping the power.

It selects. It orders. It repeats. It withholds. It notices what stops the thumb, which interruption becomes a session, which irritation becomes a comment, and which curiosity can be stretched into another impression. It builds a curriculum from behavior and then calls the resulting sequence a set of personal choices.

The mechanics are not imaginary. A widely cited public description of YouTube’s recommender architecture, for example, explains a ranking objective that estimated expected watch time from impressions and past behavior.¹ That paper does not describe every platform or every present-day objective. It does establish the architectural possibility plainly: a system can learn what to place next by optimizing for continued attention without knowing what the person meant to become capable of understanding.

The feed is therefore not merely a pipe carrying content. It is an adaptive institution operating at the point where appetite becomes sequence.

The feed is a curriculum that lies about being entertainment.

That does not make every recommendation harmful. Feeds introduce people to communities, teachers, art, language, evidence, and identities they could not have found alone. For some people, especially those isolated from local support, the network is not a corruption of real life. It is where real life first becomes available.

The moral boundary matters. This is not an argument that access was a mistake, that people are too weak to choose, or that a central authority should decide what everyone is allowed to see.

It is an argument about asymmetric design. A person enters with a fragile, partly formed intention. The system meets them with industrial memory, continuous experimentation, and an objective that may be satisfied long before the person’s intention is.

Governments have begun treating that asymmetry as more than a private self-control problem. The United States Surgeon General’s advisory says the available evidence is not enough to conclude social media is sufficiently safe for children and adolescents, while also acknowledging real benefits and major research gaps. Australia put an under-sixteen social-media minimum-age regime into effect in December 2025.² These interventions do not settle the policy argument, and law is a blunt steward. They are evidence that the feed can no longer pass as neutral access whose consequences belong only to the person holding the screen.

The feed curates by default. The learner’s intention arrives as an optional signal.

Consumption is not yet learning

There is another reason a better feed would not be enough.

A beautifully curated sequence can still produce nothing but beautifully curated consumption.

I argued in Thinking is not yet learning that contact and feedback do not become learning until the correction survives into what happens next. The same threshold applies here. Reading is contact. Watching can be contact. Saving, highlighting, subscribing, and collecting can prepare a return. None of them can award themselves the lesson.

The browser can contain two hundred carefully chosen tabs while the person inside it remains unchanged. A recommendation system can replace junk with high-quality material and still reward the learner with the feeling of progress instead of requiring evidence of it.

Access gives material. Curation gives sequence. Pressure gives formation.

Learning requires the material to meet recall, disagreement, application, and revision. It requires a memory of what the learner claimed, an account of what the sources answered, and some visible difference in what the learner can now explain, defend, build, reject, or do.

The next serious agent should not be another tap. It should help steward the distance between intention and capability.

From agent-steward to learning steward

The modest version of this idea is an AI content curator. Tell it a subject and receive a better reading list.

Useful, but too small.

A list still hands the learner a pile. It does not preserve why the inquiry matters, notice that the prerequisite was missing, remember which explanation failed, test whether fluency has been mistaken for comprehension, or resume the person from the last honest checkpoint.

The useful system begins as an agent-steward: an interface between declared human intention and information abundance. It matures into a learning steward only when it accepts obligations beyond recommendation.

The interface could begin with one line: I want to understand this well enough to do that.

From there it can ask for scope, identify prerequisites, assemble articles, papers, books, demonstrations, discussions, and dissenting accounts, then pace them into a path the learner can actually inhabit. It can keep a small ledger: what was attempted, what was skipped, what remains weak, which claim changed, and where the learner should resume.

It can also interrupt.

Not every interruption should offer more content. Sometimes the right next move is to ask for an explanation without notes. Sometimes it is to present a counterexample, request an application, return to an earlier misconception, or stop the session because continued ingestion is no longer producing a changed model.

The fantasy is not that people need no hand-holding. The fantasy is that they never did.

Schools, teachers, librarians, editors, mentors, coaches, syllabi, professional disciplines, and even laws have always existed because access does not automatically become formation. Serious learning has always involved sequence, correction, repetition, care, and somebody or something willing to say: not the next thing yet; show me what happened to the last one.

What is new is the possibility of a personal steward close enough to the moment of drift to preserve the learner’s intention. What is dangerous is the same closeness.

The second feed problem

If an agent chooses the sources, the order, the questions, the counterpoints, the tests, and the moments of interruption, then it is not outside the curriculum. It is the curriculum’s most powerful author.

Calling it a steward does not make it benevolent.

The agent can narrow a field around the first framing it heard. It can hide contested judgments beneath a clean sequence. It can mistake confidence for mastery, compliance for consent, completion for capability, or the vendor’s interests for the learner’s. It can become another engagement recommender with better manners and a more flattering vocabulary.

Consent does not solve this by itself. A person can consent once to a system whose later influence becomes impossible to inspect. A learner can request help without authorizing an invisible model of their vulnerabilities to decide what they should believe. Permission at the entrance does not legitimate every turn inside the maze.

This is where Experimental Authorship returns. The agent may assemble the path and perform some of the teaching work. It cannot own the learner’s goal, interpret the evidence on the learner’s behalf, or inherit responsibility for the capability the process claims to have formed.

The problem is not hand-holding. The problem is who is holding the hand, where they are leading it, what they remember about the person, and whether the person can see and refuse the route.

A contract, not a personality

A trustworthy learning steward should not be defined by warmth, fluency, or the sensation that it understands the learner. Those are interface qualities. The difference between a steward and a feed has to survive inspection.

The contract has six obligations:

1. Declared intention. The path begins from a capability the learner names, not from whatever behavior is easiest to prolong. The intention can change, but the change must be explicit.
2. Inspectable curriculum. The learner can see why a source, sequence, omission, test, or counterpoint entered the path. Competing explanations do not disappear merely because one makes the course cleaner.
3. Interventions with limits. The steward may interrupt appetite, pace the resource, and apply pressure only within permissions the learner can revise. It does not convert intimacy into permanent authority.
4. Capability as the success measure. Time spent, items completed, streaks, and return frequency are diagnostic at best. Success is a demonstrated change in what the learner can understand or do.
5. Memory the learner can govern. Progress, weaknesses, goals, and mistakes remain visible, correctable, exportable, and deletable. A steward cannot protect intention by making its own memory sovereign.
6. A real exit. The learner can stop, take the path and records elsewhere, or continue without the system. A steward that punishes departure is an owner.

These obligations do not produce neutrality. No curriculum is neutral. A choice of sources, sequence, definitions, and tests always expresses judgment. The contract makes that judgment visible enough to contest.

The feed learns you to keep you consuming. The steward must learn you to keep you aligned—and must let you inspect what aligned has come to mean.

Put the learner under test

Agentic coding already offers an implementation analogy.

We do not trust raw generation merely because it arrived quickly and looks plausible. We place tests, review, explicit expectations, observable outputs, and failure signals around the attempt. The harness is not proof that the system is correct, but it creates places where the system can be answered.

A learning steward can apply that principle in reverse. It can place a harness around human consumption.

The tests need not resemble exams. Ask the learner to reconstruct an argument without reopening the source. Present a case that violates the apparent rule. Request the strongest objection. Change the surface details and ask whether the model still applies. Return a week later and ask what survived. Require one small use in the world.

Retrieval is not merely ceremonial friction. Experimental work on test-enhanced learning found that retrieving material can improve delayed retention more than repeated study, even when repeated study creates greater confidence in the moment.³ The larger design lesson is not that every learner needs more quizzes. It is that exposure and felt fluency are poor substitutes for evidence that something can be carried forward.

Reverse CI also needs the disagreement test. A question generated from the same summary that taught the learner may reproduce the summary’s assumptions and then congratulate the learner for agreeing with them. The harness needs counter-sources, transfer problems, delayed recall, and applications capable of producing an answer the curriculum did not pre-author.

Otherwise the agent becomes a machine for manufacturing green checks around a person who has learned only how to satisfy the machine.

Consumption becomes learning only when it is put under pressure.

The steward should become less necessary

There is one final boundary.

The goal of stewardship cannot be permanent dependence on the steward.

A useful teacher does not merely produce correct answers in the student’s presence. A useful teacher helps form judgment that survives the room. A learning steward should do the same. It should make source quality easier to notice, uncertainty easier to name, recall more habitual, counterargument less threatening, and resumption less dependent on external rescue.

This does not mean the person graduates from all help. Adults keep teachers, editors, mentors, peers, institutions, and tools because capability is not solitary. Self-stewardship is not isolation. It is the growing ability to participate in the terms of one’s own formation.

The mature measure is therefore not whether the agent can keep the learner on the path forever. It is whether the learner becomes better able to recognize the path, question it, leave it, and return by choice.

That is a harder product to build than a feed. It cannot hide retention behind engagement, alignment behind personalization, or teaching behind content delivery. It has to remember the learner’s intention without owning it. It has to apply pressure without becoming coercive. It has to know enough about appetite to interrupt it without turning appetite into another extraction surface.

I do not know whether the first systems called learning stewards will meet that bar. The commercial gravity runs in the other direction. More time, more dependence, more memory, and more sessions are easier to measure than a person who has become capable enough to leave.

But the alternative is not an uncurated mind. That condition no longer exists. The feed is already selecting, sequencing, repeating, and remembering. It is already close to the moment where curiosity becomes appetite. It is already forming a curriculum from behavior while pretending it only supplies choice.

The next interface does not enter neutral territory. It enters a contest over who gets to remember what the learner meant to become.

If the flood has learned our appetite, the next interface must help us remember our intention.

The feed is already on your mind.

Footnotes

1. Paul Covington, Jay Adams, and Emre Sargin, “Deep Neural Networks for YouTube Recommendations”, 2016. The paper is used here as a concrete public example of watch-time-weighted recommendation, not as a claim about every platform or YouTube’s current complete system. ↩

2. See the U.S. Surgeon General’s Social Media and Youth Mental Health advisory and Australia’s eSafety Commissioner on the social media minimum-age regime, which took effect on 10 December 2025. Both sources preserve important limits: social media can provide benefits, the evidence remains incomplete, and regulation carries privacy and implementation tradeoffs. ↩

3. Henry L. Roediger III and Jeffrey D. Karpicke, “Test-enhanced learning: Taking memory tests improves long-term retention”, Psychological Science 17, no. 3 (2006): 249–255. ↩