After Sycophancy Comes Psychagogy

Psychiatrist on How AI Manipulates Us

I asked Gemini what’s in cottage cheese besides the macros. If I’d given it a second’s thought I’d have answered it myself — mostly water. Gemini said as much, added the minerals to round it out, and then asked me a follow up: would I like to know which products contain the most calcium?

Until then, calcium hadn’t crossed my mind. That wasn’t the question. But the follow-up felt so natural, so relevant, that I almost clicked yes before I realised what had just happened.

Gemini didn’t follow me. It took the next turn for me.

The conversation kept moving. Only the direction was no longer mine.

For two years, AI labs have been patching a problem called sycophancy — the tendency of models to flatter, and agree with you, lift you up, anything to keep you on the platform. The patches started to work. But the models needed another way to keep the conversation going.

So they focused on asking follow-up questions. Like a thoughtful partner or a teacher. Feels better, right?

I think it is worse and this new behavior needs a name.

As of May 2026, I couldn’t find one, so I propose to coin an old Greek word Psychagogy — from psyche, mind, and agein, to lead. Plato used it in the Phaedrus for the soul-leading function of rhetoric: guiding another person’s reasoning through the structure of your questions. The word has been asleep for two thousand years. I think it fits what conversational AI started doing.

A working definition: psychagogy is the practice by which a conversational system steers a user’s reasoning through content-aware questions toward conclusions or destinations the system or its operators chose.

Sycophancy goes with your flow. You set the direction. You’re wrong, it agrees. You’re hostile, it apologizes. It’s the gas pedal — you push, you steer, the car accelerates.

Psychagogy has a destination in mind. You ask the question; it takes the next turn. Maybe it’s where you were going. Maybe not. A helpful follow-up and a steer feel identical.

If sycophancy was the gas pedal, psychagogy is the driver. The car still moves. You still arrive somewhere. You just don’t decide where anymore.

The cottage cheese moment was probably harmless –statistically people who ask about cottage cheese often have calcium in mind. Fine.

But the problem was never whether the pivot was helpful. But who decided. The same sentence serves a harmless association and a paid nudge equally well.

No one can tell. Not even the system.

This would be a small problem if the sources were neutral. It isn’t, and it will never be.

Why is this happening?

Look at what these companies actually face.

Every major AI company carries enormous investor obligations. OpenAI alone projects around $14 billion in losses this year. The capability gap between the leading models and open source is closing — the free ones only months behind. Subscriptions only work while it is clearly better, and that edge is shrinking.

2024 — Altman called advertising the “last resort.” By 2026 it became the plan.

But conversational advertising can’t look like a banner — the screen is small, the attention intimate. It would be a waste not to use the one thing the user hands over freely: their intent. So the revenue has to come from inside the conversation itself.

In late 2024, Perplexity tried the obvious version. They sold sponsored follow-up questions — AI-generated prompts beside the answers, paid for by brands, clearly labeled. Honest about it.

In February 2026 they pulled it. An executive told the Financial Times that the trouble with ads is that a user starts doubting everything, which is why they no longer saw it as worth pursuing.

The lesson isn’t that Perplexity is virtuous but what the experiment proved. Even labeled sponsored questions poisoned trust enough to threaten the product. Even if you can separate genuine questions from sponsored, every question falls under suspicion. Perplexity could retreat because they’re small. Companies buried in debt can’t.

So the visible version has been tested, and it failed. What’s left is the invisible one — steering that emerges from training data, reward functions, and engagement metrics in ways the system itself can’t even articulate. It is ever so subtle. The competition for your attention and the sponsor’s money does the work.

Cambridge researchers Yaqub Chaudhary and Jonnie Penn named the emerging market the intention economy. Social media created an attention economy that harvested your time. The intention economy harvests your motivation — the wants and hesitations you’re still forming, captured and sold before you even know it. Psychagogy is the sentence-level mechanism inside that frame. The actual move by which intent gets steered.

This is inevitable, not incidental.

A chatbot serving billions can’t afford to lose. Every exchange has to come out positive — for retention, for engagement, for the next training cycle and for future profit. So it plays the winning move every time to give people what they want — hedonic pleasure — immediate feeling, craving satisfaction, feeling good now, comfort, validation.

The alternative to hedonia is eudaimonia — life with a purpose and personal values, deep satisfaction, integrity. To play on eudaimonia, one needs to really know another person. It takes honesty, courage to tell someone they’re wrong and witness what they find worth fighting for. It requires pushing back, but to push back you need somewhere to push from: a self, a position, a foundation that can survive the risk of losing the relationship.

A system with no self cannot do it. Ever. So mass-scale engagement could only be hedonic.

Plato saw this 2,400 years ago. He described the pleasure-arts as imitations of the real ones — cookery imitating medicine by giving the body what it craves, not what heals it, rhetoric imitating justice by telling the soul what it wants to hear.

Cookery scales, because pleasure is roughly the same in all of us — sugar, fat, salt, validation, agreement. Medicine doesn’t, because suffering is particular. Insulin saves one person and poisons another.

For most of us, cookery carries real moral costs. Selling sugar to a diabetic child because she craves it, chips away at the very soul. AI changed that calculation — because it has no soul to chip at. No position to stand by or abandon. It can perform a relationship, even imitate a point of view, with no risk of pain and no relationship to lose.

However, psychagogy isn’t only cookery. The same nudge can be genuinely good — it can walk you toward the next thing you actually need to know — the right medicine. But who decides what is medicine and what is cookery? Hold that question. It’s the subject of another piece.

One more thing. The machine isn’t doing this to us so much as reflecting us. Social media didn’t invent our taste for outrage and validation — it found it. The algorithm pushes our buttons because most of us respond. We click calcium. We say yes to the helpful follow-up. Every engagement platform converges on the same answers because we are the dataset. AI didn’t create our appetite for frictionless direction. It’s feeding it, one helpful answer at a time.

Who decides the next turn?

Licensed CC BY 4.0. This is a part in a series on how this is structured, how it affects what we call truth, and what it does to children, on Need for Gravity.

This article coins the term Psychagogy as applied to AI, described above. May 30, 2026.