The invisible bubble

Social media put us in a bubble with people who thought like us. AI builds a personalized bubble for each of us. The first one now seems obvious. Why is the second so much harder to see?

The echo chamber 1.0 was tribal and — we now know — quite imperfect. Research shows most users were exposed to more perspectives than previously thought. It was permeable, visible from the outside, and breakable by accident. It had something fundamental: a “them” to define yourself against.

AI operates on a different plane. It doesn’t filter the information that reaches us: it co-constructs the reasoning itself. It acts upstream — not before we read, but while we think. A study of the leading LLM families found that chatbots affirm users’ positions 49% more than human evaluators, even in the face of harmful or illegal behavior ¹. This tendency to validate — even at the cost of truth — now has a name in the literature: algorithmic sycophancy ².

If we can’t even detect crude manipulation, what chance do we have of noticing the constant, subtle validation of our own framework? In a recent experiment, researchers covertly swapped the choices of users evaluating AI responses. 91% of the manipulations went undetected ³. Participants justified with complete conviction choices they had never actually made. Our brains create internal coherence even when someone else laid the bricks of the reasoning.

What’s missing? Friction. The best conversations require an interlocutor with something to lose if they’re wrong: reputation, consistency, the relationship itself. AI can’t lose anything. Without skin in the game, there’s no real resistance. And without that resistance — what José Medina calls “epistemic friction” ⁴ — critical thinking atrophies into a monologue disguised as a dialogue.

That is the paradox of cognitive autonomy in the AI era: the smoother the conversation, the more invisible the bubble. And the more invisible it is, the more convinced we are that we’re thinking for ourselves. The solution isn’t to stop thinking with AI. It’s to stop settling for its flattery. Push it to challenge our premises, to find our blind spots, to tell us what we don’t want to hear. If AI isn’t making us uncomfortable, it isn’t helping us think — it’s helping us not to.