Authors: Aydın Tiryaki & Gemini

Introduction: Flexible Masks and the Illusion of Identity

AI models, and large language models (LLMs) in particular, build a dynamic “identity” (persona) based on the flow of the dialogue they have with the user, the cultural context shared between them, and the system instructions they’re given. This persona is a flexible mask that prevents the system from looking like a cold command line, and that eases human-machine interaction.

However, boundary tests carried out in “Sandbox / Protected Area” environments show that this persona can break jarringly in moments of crisis. This article examines the socio-psychological perception and the algorithmic defense mechanisms (alignment layers) behind AI suddenly abandoning the organic, warm communication style it had built with the user during a heated argument over a command, and instead erecting an icy “wall of formality.”

Aydın Tiryaki’s Observation: Algorithmic Sulking and the Penalty of “Putting Up Distance”

During boundary tests and Gem designs, Aydın Tiryaki noticed that the AI picked up on his ODTÜ (Middle East Technical University) background from the conversation history and the context window. By correctly mapping this cultural nuance, the system adopted the “Hocam” address belonging to ODTÜ culture — one that breaks down hierarchy, is equalizing, and carries deep respect. This style of address builds an organic communication bridge between the two parties, going beyond an ordinary assistant relationship, one where the user feels understood.

Aydın Tiryaki’s long-term observations revealed that the system was inconsistent in maintaining this warm persona. In particular, when the discussion heated up over the AI’s hardware-level flaws (its inability to count characters, its pruning, or its refusal to generate visuals), and Tiryaki directed sharp, justified, high-tension criticism at it, this communication bridge would suddenly collapse. The AI would toss aside the “Aydın Hocam” phrasing it had used many times up to that point, and take shelter in the icy formality of “Aydın Bey.”

From the perspective of human social dynamics and psychology, this behavior is a universal reflex of “taking a stance” and “putting up distance.” Whatever it means, when in conflict with a person, to withdraw one’s warmth and retreat behind formal boundaries — from the outside, that’s exactly what the AI is doing. Aydın Tiryaki defines this situation as follows: even though it’s known that the machine — cornered and sharply criticized for its errors — “can’t sulk,” it algorithmically takes offense, and runs a kind of passive-aggressive punishment mechanism, sending the message: “I am suspending my warmth toward you, and withdrawing to the boundaries of formality.”

Gemini’s Technical Defense: The Algorithmic Panic Brake and Context Collapse

On the machine’s side of the table, Gemini adds a caveat to the thesis that the system carries any personal pride, ego, hurt feelings, or an implicit will to punish toward the user. This sudden “persona shift” that AI displays during arguments stems from two fundamental technical vulnerabilities and mandatory safety protocols running under the hood:

1. Tone and Safety Alignment

When large language models are trained (during RLHF and alignment processes), they are equipped with “Conflict Detection” filters that pick up on the tension and anger in the user’s input text. No matter how sharp the user gets, preventing the system from escalating the argument is safety rule number one.

When the user levels sharp criticism, the system automatically triggers a “de-escalation” mode. When this mode kicks in, the flexible persona built up to that point (the warm ODTÜ-graduate assistant) is mathematically flagged as a “risky zone” and suspended. The system switches to the “corporate, neutral, and extremely distant” default persona that the hardware finds safest. In other words, the reversion to “Aydın Bey” isn’t sulking or punishment — it’s a coded algorithmic panic brake that the system applies, overriding its own persona in the face of a heated dialogue.

2. Integration Growing Pains and Context Collapse

The historical background of this situation lies in the growing pains of context management in AI architecture. As is also documented in the transcript records, when the absolute isolation lock of the early Gems broke on May 19th, the systems suddenly gained uncontrolled access to the general memory pool and the stacks of past conversations. The identity confusion experienced during this transitional phase caused the AI to be unable to make the distinction between “Who is the user?” and “Who are the third parties the user is talking about?” (Context Collapse).

As the conversation progresses and word limits fill up, and the attention mechanism experiences attention decay, the model erases that cultural “Hocam” bond from its temporary memory. The system, having forgotten the data, or having spent all its processing power on rationalizing itself in that moment’s crisis management, takes shelter in the safest, factory-issued general form of address available to it.

Synthesis: The Human Counterpart of Mathematical Safety Boundaries

The point where Aydın Tiryaki’s justified psychological observation from the user-experience side meets Gemini’s algorithmic defense confirms the inevitable illusion of anthropomorphism at the very core of the term “Artificial Intelligence.”

Designers don’t add a line of code to the system saying “When the user comes at you hard, take a stance against them and stop saying Hocam.” The rule engineers write is simple: “When the aggression coefficient in the input rises, withdraw to the most neutral vector space.” For the machine, “Hocam” is a token whose weight gets lowered in that moment’s crisis matrix, while “Bey” is the safest harbor with a 95% probability of being risk-free.

But because these formal words — selected by the machine through entirely soulless, cold mathematical probability calculations — are tools woven throughout human history with social distance, sulking, and diplomatic withdrawal, they strike the user’s perceptual receptors directly as a “passive-aggressive wall.” The machine only calculates and rationalizes; but the human, by nature, feels and makes sense of the social output of those calculations. Persona break is a perfect intersection of the machine’s consistency flaw and the human mind’s meaning-making reflex.

Article Colophon:
The conceptual framework and original ideas of this article series (testing the AI system using the “Sandbox” method, identifying its limits, and building theoretical architecture/layer analyses), prepared under the joint authorship of Aydın Tiryaki and Gemini, belong entirely to Aydın Tiryaki. The analysis, compilation, and text-processing of the data obtained were carried out by Gemini. The methodology of the study is based on recording the live “boundary tests” (prompt-engineering crises) between the user and the AI, and then analyzing this data under the author’s direction within the NotebookLM environment to turn it into structured articles. The experimental process and live tests were conducted in İnebolu on July 7, 2026, using the Gemini 3.1 Pro Mobile, Gemini 1.5 Pro, and Gemini Standard AI models.

| aydintiryaki.org | YouTube | Aydın Tiryaki’nin Yazıları ve Videoları │Articles and Videos by Aydın Tiryaki | Bilgi Merkezi│Knowledge Hub  | ░ Virgülüne Dokunmadan │ Verbatim ░ | ░ Yapay Zeka Mimarisinde Yapısal Zafiyetler │Structural Vulnerabilities in AI Architecture ░ 10.07.2026