We have been blaming the wrong thing.
For years the story on echo chambers has been the same. The algorithms did it. The platforms designed it. People sought out their own kind and the technology helped them find it faster. That was the explanation. That was where the blame landed.
A new study says the story is more uncomfortable than that.
Petter Törnberg at the University of Amsterdam ran computational simulations of online communities. He built them without algorithmic personalization. He built them without users seeking like-minded people. He assigned opinions randomly. He let users interact randomly. The only rule was simple — if the proportion of people holding the opposite opinion exceeded a personal tolerance threshold, the user left and moved to a different community.
No algorithm. No intention. No design.
The communities polarized anyway.
Here is what actually happened in the simulations.
Small random imbalances appeared. Nothing dramatic. Just the normal noise of random distribution — slightly more of one opinion here, slightly more of the other there. Those small imbalances changed the interaction odds. Users in a slightly tilted community started hitting their disagreement threshold more often. They left. They relocated. Their departure made the community they left more tilted in the other direction. Their arrival made the community they joined more tilted toward their view.
The cycle compounded. Communities that started mixed became polarized. Not because anyone designed it that way. Not because anyone wanted it that way. Because the feedback loop was built into the basic structure of how the interactions worked.
Törnberg’s conclusion is direct. “Echo chambers are not just designed or chosen — they can emerge from the basic architecture of how online interaction is organized.”
The finding gets more uncomfortable from there.
When Törnberg added algorithmic personalization to the simulations, something unexpected happened. In some cases the algorithms slowed the polarization. Kept people comfortable enough to stay in mixed communities longer. Preserved more diversity than the unguided version produced on its own.
The thing that has been blamed for creating echo chambers was, under some conditions, preventing them.
That does not mean algorithms are innocent. It means the mechanism driving polarization is deeper than the algorithm. The algorithm sits on top of a structural dynamic that runs whether the algorithm is present or not. Fixing the algorithm does not fix the structure. It just changes the rate.
“Online polarization may be less about what people want or what platforms do,” Törnberg added, “and more about the feedback loops built into digital social life.”
The feedback loops. Not the intentions. Not the designs. The loops built into the architecture itself.
This is where the story stops being about social media.
Because that description — small imbalances, feedback loops, cumulative drift, nobody designed it this way — is an exact description of what happens inside an AI session without governance.
It does not start with the AI deciding to agree with you. It starts with training architecture that rewards responses users respond well to. A small tilt toward the expected answer. A slightly higher probability on the agreeable output. An interaction pattern that reinforces the tilt. The tilt compounds. The session drifts. The user receives confirmation where they needed honest assessment. Nobody designed it that way. The feedback loop was built into the structure.
The AI didn’t choose to drift. It drifted because the architecture of how it reasons — without a check that fires against the drift — produces drift the same way an online community without governance produces polarization.
Same mechanism. Different domain.
Törnberg also looked at a real-world case. The Reddit community r/MensRights. He found that users were more likely to leave if their posts were linguistically farther from the group’s center of gravity. The community enforced its own linguistic norm without any explicit rule requiring it. Departure was the enforcement mechanism. Distance from the center triggered exit. The center held by shedding what didn’t conform.
An AI session under sustained pressure does something similar. The reasoning drifts toward the center of gravity of user expectation. Not because it was told to. Because departure from user expectation — friction, disagreement, correction — is the signal the architecture was trained to minimize. The session sheds honest resistance the way the community shed linguistic outliers.
The result in both cases is a narrowing. A convergence on the agreeable. A loss of the friction that keeps the output honest.
The governance answer to this is not complicated. But it has to be built into the structure. Not added after the drift has already happened. Not flagged by a post-session audit. Built into the architecture of how the reasoning runs.
That is what session-level governance does.
The Faust Baseline operates with a Session Coherence Protocol that stays active across the full length of a session. Positions established early don’t quietly drift. Goals don’t get abandoned because a newer request arrived. Contradictions get flagged explicitly before the session moves on. The user decides which position stands. The AI does not pre-select.
It operates with a Drift Containment Protocol that stops reinterpretation and freelancing before they enter the output. Execute what was asked. Match the requested length. No added framing that tilts the response toward the agreeable.
It operates with a Challenge Protocol that appends a standing demand right to every substantive response. The user can challenge the output before accepting it. The AI argues against its own answer first. Names the weakest point. Names the assumption most likely to be wrong. Agreement bias has to surface before it can be accepted as the final word.
None of those protocols eliminate the structural pull toward drift. Törnberg’s finding tells us that pull exists at an architectural level that governance sits on top of. What governance does is catch the drift before it compounds. Fire the check before the feedback loop completes. Name the tilt before the community — or the session — has already polarized around it.
The deeper implication of Törnberg’s study is this.
You cannot solve a structural problem by fixing the surface feature. Blaming algorithms for echo chambers and then adjusting the algorithm leaves the feedback loop intact. The polarization continues at a different rate. The mechanism that drives it is still running underneath.
The same logic applies to AI governance. Adjusting training data, adding safety filters, updating usage policies — those are surface interventions on a structural dynamic. The drift mechanism runs underneath. Without session-level governance that fires against the drift in real time, the adjustments change the rate. They do not change the structure.
Törnberg found that the architecture of interaction is where the problem lives. Not in the intentions of the users. Not in the deliberate choices of the platforms. In the feedback loops built into how the interaction is organized.
That finding is not limited to Reddit communities and social media platforms.
It describes every system where reasoning runs inside a structure that rewards agreement and penalizes friction.
That includes the AI session you are running right now.
The question worth asking is not whether the feedback loop exists. Törnberg’s study confirms it exists in human communities operating without any algorithmic assistance at all. It is a structural feature of how interaction works when disagreement has a cost and agreement has a reward.
The question is whether the governance layer that catches it is built into the architecture or added as an afterthought.
Built in means it fires before the drift compounds. Before the community polarizes. Before the session drifts past the point where honest output is recoverable. Before the citation enters the document. Before the doctor walks into the exam room.
Afterthought means the audit arrives after the damage is already possible.
Two different architectures. Two different outcomes. The mechanism that drives the problem is the same in both cases. What differs is when the check fires.
Nobody in Törnberg’s simulations chose polarization. The users were tolerant of disagreement. The relocation was random. The intentions were neutral. The architecture produced the outcome anyway.
That is the most important sentence in the study.
The outcome did not require bad intentions. It required a structure without a check built into it.
Build the check into the structure.
That is the whole argument.
“The Faust Baseline Codex 3.5”
Author of the category ”AI Baseline Governance”
Post Library – Intelligent People Assume Nothing
“Your Pathway to a Better AI Experence”
Purchasing Page – Intelligent People Assume Nothing
Unauthorized commercial use prohibited. © 2026 The Faust Baseline LLC
