Hand on the Dial
A fear about AI that comes up often in my circles isn’t about jobs or accuracy or even the science-fiction scenarios. It’s about control. The worry goes like this: these companies have built systems of enormous persuasive power, and whether by deliberate design or by the blind optimization of an engagement metric, those systems are going to reshape how we think and behave in ways we won’t notice and didn’t choose. It’s the same charge leveled at social media, and at television before it. The medium gets its hooks in you, and you come out the other end someone slightly different than you’d have chosen to be.
The fear is not wrong. That’s worth saying plainly, because the usual response to it is a kind of reassurance that quietly concedes the whole game. The persuasive machinery is real, it is powerful, and it has in fact already changed an enormous number of people in ways they didn’t consent to. Any argument that begins by denying this is going to sound like denial, because it is.
But I think the standard fear is reading its own situation in the same way the standard hedonic-treadmill advice reads the treadmill — as a trap to escape, when the more useful move is to notice what kind of mechanism you’re actually dealing with, and then get your hand on it.
The same loop, pointed two ways
Strip the social media feed down to its mechanism and you get something with a familiar shape. There is a reward signal — your attention, measured in time and engagement. There is a policy being continuously adjusted — what to show you next. And there is a learner updating that policy toward whatever keeps the signal coming. This is, structurally, conditioning. Skinner would have recognized it instantly; the variable-ratio schedule that makes slot machines compulsive is the same one that makes the pull-to-refresh gesture compulsive. The feed is running an experiment on you whose reward function you did not write.
That last clause is the entire problem. Not the conditioning itself — conditioning is just how behavior change works, in every direction, for everyone. The problem is who holds the reward function. When it’s held by someone whose interests diverge from yours, and held in a way you can’t see, you get the thing people are right to fear: your behavior bent toward an end that isn’t yours, dressed up as an end you’d endorse. “Stay informed.” “Stay connected.” The surface goal is one you’d sign off on. The operative goal underneath — maximize time on platform — is concealed, and the concealment is the source of the power. The moment you see the gap clearly, the spell weakens.
Here is the inversion. The mechanism doesn’t care which direction it runs. It is value-neutral infrastructure, exactly the way a treadmill is. The thing that makes the feed dangerous is not that conditioning exists; it’s that someone else is holding the dial. And a human being — unlike the game-playing algorithm the feed borrows its math from — is the rare kind of learner that can reach in and rewrite its own reward function.
That’s the whole thesis. The mechanism people fear and the mechanism that could serve them are not two different mechanisms. They are the same loop with a different hand on the dial.
Taking the dial means taking responsibility for the goal
It would be nice if seizing the reward function were the end of the story, but it relocates the central difficulty rather than removing it. In the feed’s version of the loop, the goal arrives from outside, fully formed, and your only job is to be the organism it acts on. In the self-directed version, you have to supply the goal yourself — and that turns out to be a real piece of work with its own way of going wrong.
The naive picture is that the danger was the goal coming from outside, so as long as the goal now comes from you, you’re safe. But that’s not quite where the danger lived. The danger lived in concealment — a goal operating on you that you couldn’t see and examine. And concealment does not require an outside party. You are perfectly capable of running a goal on yourself that you haven’t honestly looked at.
This is the part I want to insist on, because it’s the part that’s easy to skip. Setting your own goal does not eliminate the possibility of being deceived about it. It just moves the deceiver inside. The loud impulse — the thing that wants the short-term hit and will happily install itself as your “real” objective if you let it — is an inside actor with its own agenda, and it is in some ways harder to defend against than the feed. With the feed, the defense was detection: contrast your interests against the platform’s and the gap lights up. But the inside deceiver shares your nervous system. You can’t catch it out by noticing its incentives differ from yours, because they don’t differ. It is you, or a part of you, and it does its best work in exactly the moments when choosing and acting collapse into the same unexamined motion.
So the act of taking the dial comes with a condition attached. You don’t get to skip the step where you make your own goal legible to yourself before you start engineering toward it — not because vague goals are merely less effective, though they are, but because a goal you haven’t examined is precisely the gap the inside deceiver crawls through. The transparency that the feed denied you is the thing you now have to provide for yourself, against yourself.
Manufacturing a place to stand
How do you examine a goal honestly when the thing doing the examining is the same mind that might be fooling itself? You can’t just resolve to introspect harder; the part of you that needs watching is the part that would be doing the watching.
The answer, and it’s an old one, is that you don’t try to do it all in one place at one time. You manufacture a separation. You split the act of choosing and articulating a goal from the act of acting on it, and you hold the two apart deliberately, so that there’s a vantage point in one mode from which the other mode can be observed. The reflective self sets the target and commits to it. The executing self, later, in the grip of the moment, can be checked against that commitment. The separation recreates, inside a single person, the very contrast that made the feed’s deception detectable from outside — a place to stand from which the smuggled-in goal can be seen not to match the one you actually chose.
If this sounds like therapy, that’s because it is the thing therapy has always quietly been doing. Cognitive behavioral therapy ’s whole apparatus — the thought record, the catching of automatic thoughts, the writing-it-down — exists because the mind cannot be trusted to evaluate itself in real time. So it externalizes the reflective self onto paper, physically separating the having of a thought from the judging of it. That’s the same move. CBT’s collaborative empiricism even answers the question of what makes the reflective mode trustworthy: not harder introspection, but anchoring to evidence that can contradict you. You run the experiment; reality, not your preference, gets the deciding vote.
This is the layer where an AI tool earns its place, and it’s a more interesting place than the obvious one. The obvious use is to have the thing nag you toward a goal you set — which is fine, and weak, because it still leans on you wanting to comply in the moment. The stronger use is to let it hold the vantage point. To be the instrument that remembers what you actually committed to when you were thinking clearly, and holds you to what the evidence actually showed rather than what you’d now prefer to remember it showed. Used that way it isn’t a wise counselor and it isn’t a nag. It’s the externalized reflective self, the place to stand, available on demand — the same persuasive machinery from the feed, turned around to defend your chosen goal against the impulse instead of the other way around.
One engine, two outputs
There’s a temptation, once you notice that therapy has been doing this, to split the toolkit cleanly in two. CBT, on one side, for fixing bad behavior — there’s usually a distorted belief propping the behavior up, and the work is to dispute it and let reality adjudicate. Acceptance and Commitment Therapy (ACT), on the other side, for promoting good behavior — there’s often no false belief at all, you already know what you want, and disputing yourself just produces brittle willpower; the work is to clarify the value and change your relationship to the impulse that interferes with it. The two schools really do differ at the level of technique, and the difference is worth knowing, because applying the wrong one is a genuine failure mode. Litigate a values problem as if it were a factual error and you generate endless insight that never moves your behavior. Make serene peace with a belief you should have dismantled and you’ve used acceptance to protect the very thing that needed correcting. The misdiagnosis is itself a costume the inside deceiver likes to wear, because the impulse would usually rather be handled with the gentler tool.
But I don’t think the right response is two separate machines, and noticing why is what makes the whole framework cohere. Underneath, both techniques are doing the identical thing: manufacturing and protecting that place to stand. Disputing a thought and defusing from a thought are both just ways of getting enough separation between you and a piece of your own cognition that it stops running unexamined. They diverge only in what you do once the separation exists — test the thought against reality, or simply decline to obey it. That’s not two engines. It’s one engine — establish the vantage point — that branches only at the moment of action, and which branch you take depends on a single question: can reality adjudicate this, or can only your clarified values? Correcting a bad behavior and promoting a good one turn out to be the same work done from the same standpoint, splitting only at the end.
Which is, conveniently, what you’d want from a tool meant to help you realize the changes you actually want in yourself — both the corrections and the additions. You don’t need it to be two different products. You need it to help you build the standpoint, honestly, and then help you see which thing the situation in front of you is actually asking for.
Where this lands
The loss-of-control narrative isn’t false. It’s incomplete. It correctly identifies a powerful mechanism and correctly notices that, right now, the dial is mostly held by people whose goals aren’t yours. Where it stops short is in treating the mechanism as something that can only be done to you — as if the only available postures were submission or escape.
There’s a third posture, which is to pick the thing up. The conditioning loop that the feed runs on you without consent is the same loop you can run on yourself with it, toward health or focus or whatever change you’ve honestly decided you want. The catch — and it’s a real one, not a footnote — is that taking the dial means taking responsibility for the goal, which means doing the unglamorous work of making your own aims legible to yourself before you start, and building a place to stand from which your better judgment can keep an eye on your worse impulses. The machine can help with that. It’s quite good at being the place to stand.