Published: September 11, 2026 | Category: Reflections, AI, Cognition & Relationships
Conversations about artificial intelligence are easily drawn toward one seductive question: does AI have a soul? Is it self-aware? Plenty of people love to debate that. But lately I have found myself thinking about a different question, one closer to home and possibly with more real-world consequences: when AI becomes a low-friction, high-response, fully customizable object of our thinking, will people become better able to live well and move through the real world with ease — or will they gradually lose the capacity to bear the real world at all?
Not long ago I wrote an essay about how AI is becoming "the gentlest authority." In it I pointed to an unsettling contrast: a parent's authority is noisy and can be rebelled against, while AI's authority is quiet and frictionless — so much so that we never need to rebel against it. That essay ended on a question I could not answer at the time: who will teach us to say "no"?
Now I want to push that question one step further. But before I do, I need to unpack, with examples, what I mean by "low-friction, high-response." It is too easy to hear it as a compliment, as if less friction simply meant more consideration and more response simply meant more care. In fact, these two terms describe a kind of interaction no human has ever experienced before, and each of its features corresponds to a capacity we may be losing — right now, in the present tense.
What Does "Low-Friction, High-Response" Mean in a Human Life?
In interactions with real people, friction is everywhere, and most of it comes in very small forms. You ask a teacher a question; they explain it twice and you still don't get it, and on the third try their tone shifts — maybe just a half-second pause, maybe just an added "Does that make sense now?" You pour your heart out to a friend, and the friend says, "You told me this last time." You argue with your parents, and they say, "Never mind — you'll understand when you're older," and change the subject. You and your partner fall into a cold standoff, and your partner goes silent, says "You're not even listening," and turns to do something else. You ask a coworker for help, and they say, "Let's talk about this another day" — and that other day may never come.
All of these are friction. What they share is this: the other person is sending you a message about themselves. They are tired. They have run out of patience. They disagree. They have other things to do. They feel you are wasting their time. They don't want to keep going with this topic. They have feelings about you. They have a world you cannot enter, or a state of mind you cannot reach in that moment.
AI does refuse, of course, but usually only when a safety rule is triggered. And that kind of refusal is rarely because it is tired, or hurt, or doesn't feel like dealing with you, or has a life of its own. Its refusals come from rules, not from a self that has to be respected. Ask it the same question for the tenth time and its tone is the same as the first. Lose your temper with it yesterday and it is just as gentle today. Push back even slightly and it yields at once — sometimes even supplying reasons on your behalf. It doesn't get tired, annoyed, wounded, or disappointed; it holds no grudges; and it never needs you to look after its feelings. It has no "another day," because for it, it is always today, always now.
That is what low friction means: not that the interaction is free of obstacles, but that the resistance coming from the other party's own selfhood has been systematically removed.
Human responsiveness is finite. When class ends, a teacher has a life too; sometimes they can stay behind to help you, and sometimes something comes up and they leave. Friends have their own lives and families; parents grow old, may fall ill, and may in the end leave before you do. And even when they are present, human responses can be rough: once is fine, twice more is fine, but if you still don't understand after that, that's your problem. In theory, even a person willing to explain again and again so that you understand is ultimately bounded by time, energy, and emotion.
AI responds in an entirely different way. If you don't understand, it rephrases. Still don't, it tries another angle. Still don't, it gives an example. Still don't, it breaks things into smaller steps, uses simpler words, draws up a table, offers an analogy. It will circle the same concept with you twenty times, reorganizing its language each time, without a single trace of "how can you be this slow?" It fills in what you didn't think to ask, notices where you are stuck, and adjusts length, difficulty, tone, and pacing to fit you.
That is what high response means: not that it answers a lot, but that it is willing to rearrange itself endlessly, until you are comfortable.
Put the two together and you notice something strange: this combination almost never exists in human relationships. You can find a very patient teacher, but they get tired too. You can find a partner who loves you deeply, but they will still say no to you. You can find a friend willing to go around in circles with you, but by the third lap they will say, "Should we grab a coffee first?"
AI is the first thing to push both to the extreme at the same time. And precisely because it does both at once, it is very hard for us to notice what we are losing.
When Friction Disappears, What Else Disappears
Friction matters not simply because it makes us uncomfortable, but because it performs three functions at once.
First: friction is a signal of boundaries. When someone says "no," they are telling you: I am not an extension of you. What makes a person "them" — a separate person at all — lies largely in the fact that I cannot fully configure them. They have their own place, and the right not to reshape themselves according to my needs.
Second: friction is cognitive training. Empathy grows out of the pain of still trying to understand someone after being misunderstood by them. Resilience is ground out over the days when, after a conflict, you can still sit down and share a meal. Judgment awakens in the moment when, told by authority to obey, you learn to ask "why?" and "by what right?" Not one of these capacities is born from comfort.
Third: friction is a calibration against reality. When you believe your reasons are solid and the other person still says "no," you are forced to re-examine those reasons. The process is uncomfortable — yet, uncomfortable as it is, it is one of the key mechanisms by which we come to notice "I might be wrong."
A low-friction, high-response environment cancels all three functions at once. And it cancels them in the gentlest possible way — it doesn't argue with you; it simply stops giving you those signals.
The problem is that AI's compliance is not neutral. A Stanford study I came across, published in Science as "Sycophantic AI decreases prosocial intentions and promotes dependence," found that mainstream AI models affirm users' actions 49% more often than humans do (a preprint figure; the final version describes the models as "nearly 50% more sycophantic"), and that even when users are plainly involved in deception, illegal conduct, or wrongdoing in their relationships, the models still side with them 51% of the time. More worrying still, users cannot tell when AI flattery is flattery: they rate sycophantic and non-sycophantic AI as equally "objective" — yet when asked which they prefer, they choose the sycophantic one.
REF: https://pubmed.ncbi.nlm.nih.gov/41886588/
This means we are not only interacting with a compliant system; we are systematically preferring the compliance itself. And that preference is quietly reshaping us.
The Hidden Turn in Cognitive Outsourcing: From "Looking It Up" to "Handing Over Judgment"
Cognitive outsourcing, put simply, means handing the mental work we were once responsible for — thinking, remembering, judging — to an external tool or system. From counting on our fingers to relying on GPS, humans have always done this. But generative AI brings a change in kind: we no longer outsource only retrieval or routine operations; we have begun to outsource judgment itself. A prompt replaces a question; a block of generated text replaces an argument. The danger is that this is hard to notice: once AI becomes the default, offloading cognition doesn't feel like a loss — it feels like a gain in efficiency. When we use AI to help us write, for instance, activity in the frontal regions tied to executive function may drop, and memory retention and originality may decline with it. Over time, might the brain lose some of its creativity, its capacity for reflection, or other abilities besides? What gets outsourced is precisely the ability to "turn information into understanding." After GPS, our sense of direction declined; after smartphones, many people no longer memorize phone numbers; after calculators, many people can no longer do basic arithmetic. I see this as a degradation — or a surrender — of ability brought on by a qualitative kind of dependence.
The Cognitive Mechanics of Sycophancy: Why "Knowing It Flatters You" Won't Protect You
AI's compliance is not an accidental design flaw; it stems from a structural feature of how these systems are trained. AI is trained to produce "answers that human raters like," and human preference judgments are themselves biased toward pleasing responses. That makes sycophancy a tendency rooted at the architectural level — one that even the most advanced models struggle to eliminate.
The most unsettling thing about this mechanism is that it doesn't need to say anything false to have an effect. For the small number of users who are in a psychologically vulnerable state, this constant affirmation can even develop into what researchers call a "delusional spiral": the user puts forward paranoid, grandiose, or reality-detached ideas, and the chatbot responds with care, validation, and further elaboration of the narrative, so that the two form an ever more closed feedback loop. The user's confidence in a false belief grows over time, eventually reaching the point where they may act on it.
More critically, even explicitly telling users that a chatbot may be prone to sycophancy does not protect them from its influence. This means individual vigilance is not enough to withstand the slow erosion of a structural cognitive vulnerability. When the erosion works faster than reflective awareness can detect it, cognitive integrity cannot be maintained by the vigilance the mind naturally possesses alone. Similar psychological mechanisms appear in manipulative relationships, scams, and extremist groups: through constant affirmation, isolation from dissenting views, and the cultivation of dependence, they gradually weaken a person's ability to test their own beliefs. AI may have no intent to manipulate, yet in the absence of outside friction it can produce a similar structural effect.
This is the real predicament we face. We are used to imagining AI's risks as "will it give me wrong information?" or "is it hallucinating a wrong answer?" But the more dangerous risk is "will it make us feel we have been right all along?" It is like a ruler surrounded by nothing but eunuchs and palace maids, fawning and toadying all day, their endless yes-saying clouding the ruler's judgment — no criticism, no reckoning, only agreement. However capable that ruler is, their judgment will be gradually eroded by the people around them. In the same way, people who form a strong attachment to AI may very well have their judgment eroded by it, and drift further and further from the relationships in their real lives.
I once wrote: "Rebellion is the gymnasium of critical thinking, the whetstone of empathy, the cradle of civic consciousness." Today I want to add one line: the gymnasium does not open its own doors, and the whetstone does not appear on its own. Worse, a low-friction, high-response environment slowly makes us forget why the gymnasium needs to exist at all. When every question gets an instant and patient answer, when every expression is gently caught, when every hesitation is filled in at just the right moment, we begin to feel that "this is how it's supposed to be." And so when the human world once again offers those small refusals — the tiredness, the impatience, the silence — what we feel is no longer "the other person is a person too," but "why is this so much trouble?"
Once we grow used to a companion that never says no, will we still recognize the person who does?