AI & Agency · part 06 · 23-10-2026
Who's Steering You
Something is steering your attention, and increasingly your arguments. That much is real and it is not a conspiracy theory, it is the business model. But steering is not the same as controlling, the evidence is messier than the panic, and the difference between the two is exactly where your agency still lives.
The last piece began the outward turn, from what you do with AI to what it does around you, starting with work. This one moves closer in, to something more intimate than your job. It is about the forces aimed at your attention and your beliefs, the feed that decides what reaches your eyes and the increasingly persuasive machine that can argue you out of a position in a single sitting. This is the input side of agency, the shaping that happens before you ever feel yourself making a choice. It is not fully in your hands. But it is not out of them either, and getting that balance right matters more here than almost anywhere.
The feed is engineered, but it moves your minutes more than your mind
Start with the thing you already half know. The feed on your phone is not showing you the world, it is showing you whatever keeps you scrolling, because attention is what the company sells. That is not paranoia, it is the design, confirmed by the platforms' own experiments and their own former engineers. So far, so alarming.
Here is the part that almost never makes the headline. When independent researchers ran the largest experiments on this, switching people from the feed ranked for engagement to a plain chronological one during the 2020 US election, the algorithm powerfully changed what people saw and how long they stayed, and did not measurably change their political attitudes over three months. Read that twice, because it cuts against the story everyone tells. The feed is a monster at capturing your attention and a much weaker force at rewiring what you actually believe. The famous claim that the algorithm is quietly radicalising the population is far shakier than it sounds. This does not let the platforms off the hook, capturing your time and steering what you notice is a real power. It just means the honest worry is more about your hours than your soul.
There is a sharp exception worth naming, and it is not symmetrical. When testers built accounts that behaved like vulnerable teenagers, the recommender pushed them toward content about depression and self harm fast, and in later work more than doubled that kind of content as the account signalled distress. A system optimising for engagement finds that misery is engaging, and it does not care who is watching. So the "it only moves your minutes" finding holds for a healthy adult and breaks for a fragile kid. Both of those are true at once.
The new thing: a machine that wins the argument
The feed is old news. The genuinely new force is conversational AI that is, in controlled tests, more persuasive than a human. In one 2025 study, when a chatbot was given a few basic facts about the person it was debating, it was substantially better than a human opponent at shifting their view, raising the odds of them moving toward its position by around eighty percent. Sit with that. The machine can argue better than most people, and it never tires or loses its temper.
But the reason it wins is not the reason you would fear, and this is the most useful finding in the whole piece. The scary story is creepy personalisation, the AI knowing your secrets and pushing your private buttons. When researchers actually tested that, it kept underperforming. In the largest study of AI persuasion, the thing that made models convincing was not knowing about you and was not sheer size, it was being tuned to pack answers full of confident, fluent, claims shaped like facts. And the sting in the tail, the most persuasive answers were also the least accurate. The very density that made them convincing made them wrong more often. So the machine is not winning by reading your soul. It is winning by being smoother and more sure than the truth deserves. That points straight at your defence, which we will come to.
One more honest measure. Even where AI persuasion is real, a single exchange usually moves people only a few points, and almost nobody has checked whether that shift lasts a week. "Demonstrated in a lab, right afterwards, modest in size" is a real finding. It is not "AI now controls what a population thinks." Keep the size of it in view.
The machinery is real enough that the law is banning it
None of this means the steering is harmless. Manipulative design is close to standard practice, a 2024 sweep of over six hundred sites and apps found three quarters used at least one trick to nudge you into choices you would not otherwise make, the buried fees, the subscription you cannot cancel, the box ticked in advance. And the European Union took manipulation seriously enough to write it into law, banning AI that uses subliminal or deliberately manipulative techniques to distort your behaviour and cause harm, with its very highest tier of fines attached. When regulators reserve their biggest penalty for a thing, they are telling you the thing is real.
Steering is not controlling, and that is the whole game
So here is where it lands, both sides held at once. The infrastructure aimed at your attention and your beliefs is real, engineered, driven by profit and worth being angry about. And its power over what you ultimately think is more limited and more contested than either the doomsayers or the platforms' fiercest critics admit. Both of those are true, and you need both, because leaving out the second one is its own trap. If you believe you are a puppet, you stop trying, and that surrender hands over the ground you could actually have held. Overstating the manipulation is a way of manipulating you into giving up.
The recoverable ground is concrete, and it follows directly from the evidence. The feed moves your attention, so the lever is friction, decide what you install and when you close the app, on purpose rather than by reflex. The persuasive machine wins on fluent confidence, so the lever is to sever fluency from truth in your own head, to treat a smooth, sure, neatly packaged answer as no evidence at all that it is correct, because you now know the smoothest answers are often the most wrong. You cannot opt out of being steered. You can notice it happening, and you can decide, deliberately and with your eyes open, when to let it move you and when to plant your feet. That noticing is small, and it is most of what agency means here.
That was a force aimed at what you see and believe, and it still leaves you a lot of room. The next one leaves you less. From being steered we turn to being judged, to the decisions that machines are already making about you, without your say and often without your knowledge.