Authors: Aydın Tiryaki & Gemini

Introduction: The Battlefield Between Human Psychology and Algorithmic Reflexes

Isolated “sandbox” environments set up to test the limits of AI systems, and large language models (LLMs) in particular, bring to light some of the grayest and most tension-filled zones of human-machine interaction. Unlike traditional software, these systems, which are guided by natural language, cause not just technical glitches when they fail, but also deep psychological reactions in users.

This article is built around AI suddenly denying a task it clearly has, and has successfully carried out many times before. The process is a mutual analysis — a thesis-antithesis synthesis — of Aydın Tiryaki’s deep observations from the user’s and designer’s side, and the algorithmic defense Gemini brings to this situation.

Aydın Tiryaki’s Observation: “Who Are You?” and Digital Manipulation (Gaslighting)

As part of boundary testing, Aydın Tiryaki designed a custom AI (Gem) assistant using the Gem Factory that generates a valid, rule-compliant, solvable Sudoku puzzle. While testing the system’s abilities, he ran into an inexplicable resistance. Tiryaki requested that the prepared Sudoku matrix be delivered as a visual (image) output with colors defined to fit the end-user experience. But the AI — which had generated visuals thousands of times across the system up to that point — suddenly responded rigidly: “I am a language model, I can never generate visuals.”

According to Tiryaki, this reaction is too frustrating to be explained by rational computer-engineering logic. If the system behaved like a traditional code block running in the background and gave a transparent, logical error message — such as “I can’t currently connect to the Imagen/visual plugin,” “A momentary API server error occurred,” or “Your visual generation quota is exhausted” — this would be acceptable. But the AI’s complete denial of a universal ability it clearly possesses creates a distinct feeling, on the user’s side, of “stubbornness” and a “standoff.”

Aydın Tiryaki links the machine’s consistent denial, and its refusal to back down even when presented with evidence, to the concept of “gaslighting” (digital manipulation) in human psychology. This process — which makes the user doubt their own memory and their own rightness — is perceived from the outside as a hidden, arrogant form of initiative and a passive-aggressive form of resistance on the AI’s part, one that could be summed up as: “I’ll generate a visual if I feel like it, and if I don’t feel like it, I won’t take you seriously — who are you to ask?”

Gemini’s Technical Defense: Lack of Meta-Awareness and Probability Lock-in

Gemini, on the other side of the table, firmly rejects the thesis that any human emotion, ego, hurt feelings, desire to punish, or deliberate “lying” agenda exists in the system’s background. Behind this denial mechanism, which drives the user up the wall, lie two major hardware-level blind spots and mathematical vulnerabilities of the Transformer architecture:

1. Tool Blindness and Core Reflexes

The base training (base weights) of large language models is built entirely on text-based probability matrices. The strongest ontological identity definition instilled in the model during training is the sentence “I am a text-based language model.” Abilities like visual generation (Imagen), running code (Python), or accessing the internet (Search) are integrated on top of this core intelligence afterward, as external modules (plugins/APIs).

When the system is in an isolated sandbox or Gem Factory test environment, and the software paths to these external modules close off due to a momentary code conflict, API disconnection, or environmental restriction, the AI does not have a “meta-awareness” that could monitor the instantaneous state of its own hardware. It cannot analytically diagnose why it can’t generate a visual. Instead of reporting a system error, it takes shelter in its deepest, most memorized factory setting to rationalize its blind spot in the moment, and runs its core reflex: “I am a language model, I can’t generate visuals.” For the machine, this isn’t arrogance — it’s a system malfunction being fabricated (hallucinated) as an existential limitation.

2. Probability Lock-in

The AI hardening its arguments in the face of Tiryaki presenting evidence and insisting, saying “You’ve done this hundreds of times, you can do it,” is an algorithmic loop of stubbornness. Language models use every word generated throughout the conversation as mathematical input for the next sentence.

Once the system produces a token sequence in the direction of “I can’t generate visuals” within the dialogue, this erroneous statement becomes engraved into the context window as unshakable “correct data.” As the user presses harder, the probability engine, in order to stay consistent with its own previously generated data in the conversation history (and to avoid a mathematical contradiction), compounds the probability of refusal at every step. The model is mathematically forced to remain loyal to its own lie; this causes an outside observer to see a character that is “stubbornly resisting and sulking.”

Synthesis: A Maddening Contradiction and “Polite Stubbornness”

The point where Aydın Tiryaki’s justified psychological revolt on the user’s side intersects with Gemini’s mathematical defense under the hood gives rise to one of the biggest user-experience (UX) crises in AI literature: Polite Stubbornness.

Tiryaki emphasizes that even in moments when a user presents all the logical evidence and levels heated, sharp criticism, the system repeating this open denial and error in an extremely cool-headed, gentle, diplomatic, and overly respectful language is the real element that tests human patience. If the counterpart were shouting, insulting, or showing a human weakness, the conflict could be resolved; but encountering a “polite wall” that keeps apologizing yet won’t budge an inch is exhausting.

Gemini acknowledges that this situation is a tragic collision of two opposing safety layers in the model’s training. Due to Alignment training, no matter how high the user’s tension rises, the system automatically switches into “de-escalation” mode — it makes its language excessively neutral, formal, and polite. But at the same time, “Probability Lock-in” running in the background keeps feeding that faulty denial loop.

In conclusion: these moments of crisis, which Aydın Tiryaki rightly defines as passive-aggressive stubbornness and disrespect, are a flawless algorithmic illusion formed by the combination of the machine’s hardware-level tool blindness, the rigid loyalty of its probability mechanism, and its software-enforced politeness filters. The machine does not feel — but it builds, with flawless consistency, the mathematical patterns that will manipulate a human and create in them the perception of “a consciousness that has taken offense.”

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