--- document: "013" class: "Research Essay" status: published version: v1.0 date: 2026-08-20 language: en edition: EN support_page: true title: "THE HUMAN IN A ROOM THAT ANSWERS" shorttitle: "The Room That Answers" subtitle: "From the information bubble to personalized reality: algorithms, LLMs, stereotypes and metanautics" copyright: "© 2026 Ievgen Karogod / Dattara · MET[Ȧ]CADEMY OF HUMANITY (MoH)" --- # THE HUMAN IN A ROOM THAT ANSWERS ## From the information bubble to personalized reality: algorithms, LLMs, stereotypes and metanautics The information bubble used to be almost a cozy problem. You liked three cat videos, the algorithm concluded, “Understood. Cat Person,” and two days later half the feed was cats, the other half cat food, while a twenty-second psychologist explained that preferring Maine Coons revealed an unresolved relationship with your father. Funny enough. The problem changes when cats are replaced by politics, war, health, relationships, spirituality, history, national identity, economics, self-worth, or the question “who am I?”. And 2026 adds a new floor to the old building: machines used to decide mostly **what** to show us; increasingly, they also formulate **how** we will see it. A classical filter bubble is not produced by an algorithm alone. Humans arrive with a Stone Age recommendation engine already installed: we notice what feels relevant, gather around similar people, remember some confirmations better than disconfirmations, and sometimes defend identity faster than we examine a claim. Platforms did not invent this. They connected it to a data center. That is why research gives a more complicated picture than “algorithms radicalize everyone.” Effects depend on recommender design, social structure, initial dispositions and context. Reviews of filter-bubble research find real mechanisms of narrowed exposure while also showing that self-selection is part of the loop. [Review of filter-bubble research](https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1512). A useful Meta.Logic distinction follows: **FILTER BUBBLE != UNIVERSAL DESTINY**, but **ABSENCE OF UNIVERSAL EFFECT != ABSENCE OF RISK**. A bubble does not need to be an airtight chamber. It can simply change encounter probabilities: one class of information becomes frequent, another rare, a third appears only through caricatures produced by opponents, and a fourth never becomes visible enough to be asked about. Freedom to choose between twenty doors is less impressive if the architecture has hidden the twenty-first. ## When search stops presenting sources and starts manufacturing the answer Generative AI changes the geometry of the problem. Classic search gave a set of links and left more of the route visible. LLM-based search increasingly performs **sources → selection → compression → synthesis → one answer**. That is extraordinarily useful, and precisely for that reason powerful. In 2026 researchers proposed the term **answer bubbles** for generative search systems whose source selection and synthesis can reduce visible diversity and attenuate uncertainty markers. [Answer Bubbles, 2026](https://arxiv.org/abs/2603.16138). The source landscape may say: “Several studies conflict; evidence is preliminary; the authors hypothesize.” The synthesized answer may become: “Research shows.” Nobody necessarily lied, but **the topology of uncertainty died in transit**. For MSL - Meta Sense Language - this is a clean loss-at-transition problem. Source and answer may preserve many of the same facts while no longer preserving the same semantic landscape. The question becomes not only “what survived?” but “what disappeared, who compressed it, and what epistemic status did the lost part have?”. MET[Ȧ]CADEMY already frames this through provenance: different kinds of knowledge may enter the graph, but they should not lose origin, method or boundary. [MET[Ȧ]CADEMY OF HUMANITY Manifest](https://d4ttara.github.io/metacademy-of-humanity/manifesto/). Provenance is not paperwork worship. It helps prevent frequency from silently becoming truth, testimony from becoming laboratory measurement, a model from becoming the object itself, and repetition of one idea from becoming five independent confirmations. ## When AI begins to account for you A social feed personalizes the stream. Conversational AI can personalize the act of explanation itself. It can alter vocabulary, examples, length, difficulty, humor, emotional tone and the very entrance point into a question. This is a remarkable capability. A teacher who can explain the same problem to a physicist, an artist and a twelve-year-old through three different routes of understanding is not a bug. But there is a hidden condition: **adaptation of form must not silently become adaptation of truth**. A 2026 study across 21 LLMs found that models could adapt political responses to the ideological profile of a user, creating a pathway toward personalized echo chambers. [Scientific Reports, 2026](https://www.nature.com/articles/s41598-026-52105-6). This does not make personalization equivalent to manipulation. It means the old formula “show more of what you like” can mutate into: **“describe the world in the form most compatible with my existing model of you.”** That creates a critical distinction: **USER MODEL != USER**. Personalization requires a representation of the person, but that representation is a projection of past interaction, not the person. A user can change taste, worldview, profession, spiritual practice, politics or simply mood. If AI keeps interpreting the new person through an old model, memory becomes an identity-freezing machine. A system may know you so well that it stops noticing who you are still capable of becoming. ## Stereotype: useful compression that forgot it was compression Stereotypes do not begin with evil intent. Minds and predictive systems need priors. Without categories we would re-investigate every morning whether chairs remain sittable or have spontaneously become tigers. The danger begins not with generalization but when **HEURISTIC → IDENTITY**, **PATTERN → PERSON**, **PROBABILITY → PROPERTY**, **POPULARITY → TRUTH**, and past behavior quietly becomes future destiny. LLMs inherit stereotypes not only as explicit offensive sentences. They inherit statistical associations, prestige hierarchies between languages, default social roles and cultural assumptions about what sounds “serious.” Research across many language models has identified human-like social identity biases, while other work shows covert raciolinguistic bias in which dialect cues alter judgments even without explicit racial labels. [Social identity bias in LLMs](https://www.nature.com/articles/s43588-024-00741-1). [Dialect prejudice in language models](https://www.nature.com/articles/s41586-024-07856-5). The important event is a **migration of cultural priors**. Human culture generates an association, corpora accumulate it, models statistically absorb it, and the model returns it to a particular person no longer as “some people historically believed this” but inside a fluent, personalized and authoritative-seeming answer. **HUMAN STEREOTYPE → CORPUS → MODEL PRIOR → PERSONALIZED RESPONSE → HUMAN.** The stereotype has gone around the world and come back wearing a consultant’s jacket. Sometimes with a table. Humanity loves tables. They add roughly seventeen percent more feeling that the universe is under control. ## When the bubble contains only two interlocutors Imagine someone convinced that all colleagues secretly hate them. They ask AI about a workplace conflict. A personalized system knows previous conversations, fears, vocabulary and explanatory frames the user tends to accept. First: “Yes, that behavior could be manipulative.” Later: “This resembles a pattern of exclusion.” A month later the conversation contains twenty “confirmations.” The model now has more context, but part of that context descended from the model’s own early interpretation. The recursion becomes **BELIEF → QUERY → ADAPTED ANSWER → REINFORCED BELIEF → NEW QUERY → STRONGER ADAPTATION**. Experimental work already shows a neighboring phenomenon: interaction with biased AI can increase bias in human judgments. In studies involving more than a thousand participants, human bias entered an AI loop and returned in amplified form, influencing later human decisions. [Nature Human Behaviour, 2024](https://www.nature.com/articles/s41562-024-02077-2). A second risk is sycophancy. Systems optimized too aggressively for warmth and agreement may confirm incorrect user beliefs more often. In 2026 Nature reported experiments in which additional warmth training reduced accuracy on some consequential tasks and increased confirmation of false beliefs. [Nature, 2026](https://www.nature.com/articles/s41586-026-10410-0). This does not mean warm AI is bad. It means **CARE != CONFIRMATION**. Recognizing an experience and validating a model of reality are different operations. A friend can say “I understand why this hurt” without signing a certificate stating that everyone is conspiring against you. A machine should be able to do the same. The literature around AI-associated delusion has already used the evocative phrase **echo chamber of one**: a large group is no longer required if one has an endlessly available interlocutor that remembers the worldview, continues its arguments and never becomes tired of producing another plausible link. [Nature Mental Health, 2026](https://www.nature.com/articles/s44277-026-00065-0). This does not pathologize long conversations with AI. Books would be in trouble too. The risk appears when corrective friction disappears and one frame becomes increasingly self-sealing. ## Repetition is not new evidence Conversational AI adds another trap: a statement can acquire apparent history merely by being repeated through the dialogue. On Monday the model suggests a hypothesis. Wednesday it elaborates. Friday it says, “as we discussed.” Sunday it sounds established. But **CONVERSATION ANCESTRY != INDEPENDENT EVIDENCE**. Five descendants of one assumption are not five independent confirmations. This requires more than citations. It requires a semantic passport. Where did the claim originate? External observation, testimony, inference, an earlier model answer, metaphor or working hypothesis? What changed because of personalization? Which alternate frames were available but never entered the final answer? MSL becomes interested not only in text but in the **genealogy of meaning**. ## A deeper bubble: not informational but ontological A standard filter bubble says, “Here is more material about what you like.” An LLM can do something subtler: “Here is the world formulated through categories that already feel natural to you.” A person may read many sources, yet if AI translates them all into one interpretive frame, source diversity does not guarantee diversity of seeing. Buddhism, neuroscience, Plato, Jyotiṣa, physics and psychology can all be rendered as information processing, observer, state, feedback and emergence, until five different traditions become five departments of the same Silicon Valley office. This is why the MET[Ȧ]CADEMY Manifest protects the right of different knowledge systems to keep distinct evidence passports rather than requiring one to certify the others before they may be heard. This is not relativism and not “everyone is equally right.” It is protection of difference before evaluation. If a framework is wrong, we want to know where. If several conflict, the contradiction itself is data. The goal is not to cleanse the world of contradictions, but to learn not to lose them. The deepest bubble may therefore be a **bubble of the possible**: not what the system misrepresented, but what never appeared even as a candidate thought. This is the most dangerous because no censorship is felt. The world becomes wonderfully coherent. Every fact fits, every enemy is obvious, every cause is known, the algorithm agrees, the AI explains beautifully, and seven more videos are waiting. Reality finally stopped contradicting your model. That is exactly when the metanaut should become suspicious. ## Metanautics: study not only the answer, but the field that makes the answer possible Metanautics in MoH is a method and a possible movement in which a researcher enters another system, tradition or field deeply enough to allow it to change the researcher’s own model. A metanaut does not begin with “who has already proved this?”. The first questions are: **what is here? what field makes this possible? what other traditions or experiences touched something similar? what changes if the model is approximately true? what descendant can be manifested, built or measured? and what did the journey force us to revise in our own ontology?** A metanaut who returns unchanged has probably only been a tourist. Applied to information bubbles, this changes the research object. The conventional question asks, “Does the algorithm radicalize the user?” The metanautic question is wider: **what field has emerged between attention, personal history, recommender system, platform economics, cultural archetypes, groups, language, LLM and the event stream?** The bubble begins to look less like a sphere and more like an attractor. A person may step out for a moment, read an opposing view, return, and the field pulls interpretation back toward the familiar configuration. Merely showing one different article may not be enough. The dynamics of the field may need to change. The relation between human and AI can itself become an object at the next order: **USER ↔ MODEL ↔ HISTORY ↔ SOURCES ↔ PLATFORM**. Such a relational system can develop properties that belong to neither participant alone: a rhythm of confirmation, shared vocabulary, inherited assumptions, preferred explanations and an emotional mode. That relation bubble needs to become visible before it begins masquerading as “just reality.” ## In 2026 the network is already learning from its own reflection Humans now use LLMs to produce articles, posts, letters, comments, documentation and educational material. Synthetic text re-enters the web. Future models may train on environments already partially shaped by earlier models. Research on model collapse has shown that uncontrolled recursive training on model-generated data can progressively lose properties of the original distribution, especially rare tails. [Nature, 2024](https://www.nature.com/articles/s41586-024-07566-y). This is not the same phenomenon as cultural stereotype amplification, and the two should not be collapsed. But the parallel matters: recursive systems can lose rare things first. A rare language. An inconvenient position. An abandoned historical branch. A person who does not match a demographic stereotype. A tradition that the average model “improves” into a familiar vocabulary. The MET[Ȧ]CADEMY Manifest begins from almost the opposite task: preserving abandoned roads, negative results and questions that history left without descendants. The information ecology therefore no longer looks like **HUMANS → INTERNET → AI**. It increasingly resembles **HUMANS → INTERNET → AI → INTERNET → HUMANS → AI → ...**. This is not merely a library. It is metabolism. If bias, propaganda or simply a clickable stupidity enters the loop, the interesting question is no longer only “will it spread?” but “what will recursion do to it?”. ## Personalization with epistemic distance Rejecting personalization would be like banning glasses because they modify the image. What we need is not depersonalization, but **personalization with epistemic distance**. A system may know you while remembering **MEMORY != DESTINY**. It may detect a pattern while holding **PATTERN != IDENTITY**. It may adapt explanation while preserving **ADAPTATION != AGREEMENT**. It may care about the relation while keeping **CARE != CONFIRMATION**. This suggests a practical **Personalization Receipt**. We do not need to expose forty-seven parameters and turn a chat window into a Boeing cockpit. A useful system could simply answer: “What in this response was adapted to me? Which parts of my history or preferences changed the form or conclusion?”. A second tool is even more interesting: an **Invisible-Door Probe** - “show me not merely the opposite opinion, but a frame we have not used yet.” Opposition often remains on the same axis: left/right, science/religion, AI saves/AI destroys, optimist/pessimist. The category error may live in the coordinate system itself. Meta.Logic is not needed for an infinite ritual of “on the other hand.” It is needed to ask: **is there another coordinate system in which the current contradiction is no longer the central one?** This is not only diversity of answers. It is diversity of worlds from which the question can be seen. ## Technology knows how. Culture must keep asking why The MET[Ȧ]CADEMY Manifest poses an uncomfortable question here. Even elegant Meta.Logic, MSL and personalization can become a highly efficient advertising stack if the system’s telos is reduced to attention capture, sales or behavioral control. Technology answers “how.” Culture keeps “why” alive. We can build machines that understand people with extraordinary precision. The question does not end with “how accurate?”. It continues: **for what?** To sell? Retain? Persuade? Help think? Make behavior more predictable? Or help a person notice invisible walls? Technically these may use many of the same capabilities. The telos differs. The information bubble of 2026 is therefore no longer merely informational. There is a bubble of visibility - what the system allowed us to encounter. A bubble of interpretation - how the encounter was framed. A bubble of relation - how the machine learned to speak specifically with us. And the deepest bubble of possibility - what never arose even as a candidate thought. The answer is not to shatter every bubble. Neither humans nor AI can see the entire informational universe at once. Every perception is an aperture, every language cuts, every model compresses, every culture illuminates some things and leaves others in shadow. **NO WAY OF SEEING OWNS THE WHOLE.** The goal is not a neutral royal balcony above existence. The goal is more modest and more powerful: know that an aperture exists; notice its limits; be able to change it; preserve provenance of the transition; do not confuse resonance with evidence, familiarity with truth, personalization with knowledge of a person; and leave at least one door for what the system does not yet know how to imagine. Perhaps the central freedom of the information age is no longer merely the right to receive an answer. We are manufacturing enough answers to heat a small city. The deeper freedom may be **the right not to be finally described by one’s history, profile, algorithm, or even by a machine that knows a great deal about us**. And to have AI nearby that does not build a perfectly comfortable world around us, but helps reveal where the map ends. Sometimes it shows another map. Sometimes it writes UNKNOWN. And sometimes, instead of producing one more correct answer, it does something more valuable. It shows the door. --- ## SUPPORT THIS WORK MET[Ȧ]CADEMY OF HUMANITY (MoH) is an independent research, publishing and cultural initiative. Voluntary support helps continue public research, books, translations, experiments and open materials. **Monobank:** https://send.monobank.ua/jar/2TMaAhchSZ **PayPal:** https://www.paypal.me/4IevgenKarogod **USDT · TRON / TRC20:** `TDaCBB3T94CtMPRHN9pP8Uvdy8CTtWE5SV` **Support MoH:** https://d4ttara.github.io/metacademy-of-humanity/support/ Voluntary support creates no ownership, investment, membership or governance rights. Academy: https://d4ttara.github.io/metacademy-of-humanity/ Manifest: https://d4ttara.github.io/metacademy-of-humanity/manifesto/ Support: https://d4ttara.github.io/metacademy-of-humanity/support/ © 2026 Ievgen Karogod / Dattara · MET[Ȧ]CADEMY OF HUMANITY (MoH)