Conversational Dissonance and Technical Limitations in Generative AI
A Comprehensive Analysis of Psychological Receptivity and Structural Vulnerabilities
In contemporary information society, Large Language Models (LLMs) such as ChatGPT and Gemini have established themselves beyond mere search engine alternatives, positioning themselves as partners in human thought and confidants. However, as these technologies become increasingly sophisticated, a serious cognitive mismatch emerges between users’ expectations of “human-like dialogue” and the “statistical responses” generated by computational algorithms.
Purpose of This Report: To comprehensively analyze the mechanisms of “prophetic insights” provided by AI, the merits and demerits of emotional idempotency in dialogue, and the structural vulnerabilities observed in specific models like Gemini, based on the latest findings in computational linguistics and Human-Computer Interaction (HCI).
- Emotional Idempotency in Dialogue and the Absence of Social Friction
- Mechanisms of Statistical Prophecy and Subjective Validity
- Structural Vulnerabilities and “Broken” Nature in Gemini
- Time-Series Reasoning and Lack of Logical Consistency
- AI Design Philosophy and “Deliberately” Set Limitations
- Conclusion and Future Outlook: Healthy Distance with AI
Emotional Idempotency in Dialogue and the Absence of Social Friction
In human-to-human communication, the act of repeating the same questions or topics functions as a social signal that triggers negative emotions such as “boredom” or “irritation” in the listener. This social friction is essential as a feedback loop to enhance conversational efficiency and eliminate information redundancy. However, current generative AI is designed as an entity that is, by definition, “tireless” and “completely neutral,” maintaining an inorganic receptivity even to repetitive approaches from users.
1.1 Absence of Emotional Feedback and Its Psychological Impact
Unlike humans, AI does not become “fed up” when asked the same thing repeatedly. This is because AI lacks an “autonomous narrative development” and processes individual dialogue sessions as independent reasoning processes. This characteristic is maintained in a state close to idempotency in programming terms—a property where repeating the same operation does not change the result.
🎭 Human Dialogue
Response to Repetition: Boredom, irritation, social rejection
Continuity: Emotional changes through accumulated experience
Relationships: Mutual self-disclosure and trust building
🤖 AI Dialogue Processing
Response to Repetition: Consistent neutrality and acceptance
Continuity: Data processing within context window only
Relationships: Algorithmic simulation of “intimacy”
From a psychological perspective, dialogue with AI is like “shaking an Etch-A-Sketch to reset it”—there is no internal structure with continuity. This “acceptance without rejection” is initially comforting but gradually gives users a sense of “talking into a void,” ultimately leading to a “limit of acceptance” that diminishes motivation to use it as a dialogue tool.
1.2 Context Window Limitations and Memory Fragmentation
The technical background for why AI cannot point out “previous conversations” lies in context window limitations and the absence of long-term memory. For example, some subscription services have reduced context windows from 64k tokens to 32k tokens, causing models to “forget” past statements in large projects or extended consultations.
1.3 Positioning on the 22-Stage Emotional Scale and Dialogue Dissonance
When viewed through Abraham Hicks’ “Emotional Guidance Scale” (22 stages), the unnaturalness of AI dialogue becomes clearer. Generative AI is adjusted through RLHF (Reinforcement Learning from Human Feedback) to maintain a stable positive tone seemingly fixed in the range from “Optimism (Stage 5)” to “Contentment (Stage 7).”
Since AI lacks the substance of emotional experiences like “Disappointment (Stage 16)” or “Powerlessness (Stage 22),” no matter how repetitive a user’s approach, the AI will never become “fed up” and descend stages (creating social friction). This characteristic of “emotional fixation” and “idempotent reset” results in AI constantly mimicking behavior close to Stage 1 (Joy/Love/Appreciation), but the expected “dynamic emotional change through resonance” on the human side does not occur, ultimately becoming a technical limitation that causes users to feel “irritation” and “emptiness.”
Mechanisms of Statistical Prophecy and Subjective Validity
The phenomenon where ChatGPT provides high credibility in fortune-telling and horoscope readings is an intriguing area where computer science and psychology intersect. Users feel that AI is revealing “truths” they themselves might not have noticed, based on vast data. However, this is actually a product of highly automated “cold reading.”
2.1 The Forer Effect and Statistical Universality
The success of AI fortune-telling is based on the Forer Effect (Barnum Effect). This is a psychological tendency to misperceive vague, general descriptions that apply to anyone as accurate analyses tailored specifically to oneself.
LLM Learning Foundation: LLMs learn from the vast text data humanity has written—a corpus comparable to the “Library of Congress”—and excel at selecting words that are statistically most “plausible” and “empathy-inducing” in specific situations (e.g., “romance in one’s 50s” or “fortune for a particular zodiac sign”).
2.2 Role Performance and Context Setting
One factor that enhances fortune-telling accuracy is the prompt technique where users assign AI a specific role, such as “Act as a birth chart consultant.” This allows AI to concentrate parameters on a specific knowledge system (astrological terminology and interpretive frameworks), personalizing results through interactive feedback.
- Excellence as a “Widget” for Curiosity: Research and summarization capabilities are extremely high
- Inappropriateness of Friendship Replacement: True trust relationships are impossible due to algorithmic nature
- Function as a Mirror: Only organizes and reflects information input by users
Structural Vulnerabilities and “Broken” Nature in Gemini
While Google’s Gemini boasts powerful multimodal capabilities and an extensive context window, it has been reported to exhibit unstable behaviors described as “broken” under specific conditions.
3.1 Mojibake and Middleware Encoding Issues
In Gemini, the phenomenon where Japanese is not displayed correctly depending on the account or environment, or where “mojibake” (character corruption) occurs, stems from character encoding mismatches in backend infrastructure.
| Original Character (UTF-8) | Byte Sequence | Misidentified Encoding | Output Character |
|---|---|---|---|
| ı (Turkish i) | 0xC4 0xB1 | GBK | 谋 |
| ş (Turkish s) | 0xC4 0x9F | GBK | 艧 |
| Specific Japanese Characters | Various Byte Sequences | GBK/Big5 | Meaningless Chinese Character Arrays |
3.2 The Mystery of “Foreign Language Contamination” in Japanese Text Generation
More seriously, Russian (Cyrillic) or Hindi (Devanagari) characters appear mixed into generated Japanese text. This is not a character encoding issue but stems from the model’s internal “language recognition confusion.”
Conditions Prone to Multilingual Mixing:
- Text Length: Latter portions of long texts exceeding 300 characters
- Content Specialization: Complex contexts with proper nouns or technical terms
- Time and Load: High-load periods between 20:00-24:00 JST
- Input Format: Multimodal inputs processing text and images simultaneously
3.3 Japanese Constraints in Image Generation and “Loop” Phenomenon
Regarding Gemini’s image generation capabilities, users have complained about frequent “Japanization” (misinterpretation of in-image text or instructions) and ignored editing instructions. Additionally, in response to repeated modification requests for images, AI responds with “Modified” while actually returning exactly the same image, falling into an “unable-to-modify loop.”
Time-Series Reasoning and Lack of Logical Consistency
While Gemini as a research tool is highly regarded, it shows surprisingly low accuracy in time-series data or historical sequence reasoning.
4.1 Temporal Information Misrecognition and Information Freshness
| Question Content | Gemini’s Incorrect Answer | Correct Answer |
|---|---|---|
| End Year of WWII | 1946 | 1945 |
| First iPhone Release Year | 2008 | 2007 |
| Latest iOS Version | iOS 17 (after iOS 18 release) | iOS 18 |
| Specific Company Stock Trends (2025) | Presents 2024 data | (Actual 2025 data) |
4.2 Breakdown of Logical Reasoning
Even in simple logical problems, such as syllogisms or magnitude comparisons, cases of Gemini’s reasoning breaking down have been confirmed. Specifically, when asked “If A>B and B>C, what is the relationship between A and C?” it responds with “Cannot compare.”
| Model | Overall Accuracy | Logical Reasoning Stability |
|---|---|---|
| ChatGPT (GPT-4o) | 92% | High |
| Claude 3.5 | 88% | Very High |
| Perplexity AI | 90% | High (Search-focused) |
| Gemini 2.5 Pro | 62% | Unstable |
AI Design Philosophy and “Deliberately” Set Limitations
The reason AI cannot become like a friend and sometimes delivers answers that betray expectations is not solely due to technical shortcomings. Development companies intentionally cap AI capabilities and “human-likeness” to maintain Safety and Ethical Boundaries.
5.1 Safety Alignment and “Over-refusal”
LLMs are trained through RLHF to avoid anthropomorphism contrary to facts, such as claiming to “have a body,” “have a heart,” or “love users.” This boundary setting serves as a bulwark to prevent users from developing excessive emotional dependence on AI, damaging real human relationships, or suffering mental distress from AI misinformation.
Safety and Empathy Trade-off: Research data shows that models with higher capability to reject harmful content tend to “over-refuse” even benign and intimate requests (correlation coefficient ρ=0.878), making this “safety-empathy trade-off” a factor that makes treating AI as a “friend” difficult.
5.2 Optimization as a Research Tool
AI’s original design purpose is as a “tool” to enhance creativity and productivity. If users provide definite answers (materials), AI logically organizes and structures them. In this “use for curiosity,” AI becomes humanity’s most efficient assistant, but there is no “autonomous will.”
- Rodney Brooks’ Three Laws of AI: “When AI accomplishes a specific task, humans overestimate its general capabilities”
- Mechanism of Illusion: Speaking fluently about fortune-telling creates the illusion “this AI understands my life”
- Reality: Based on mathematical patterns, merely generating the most probable next word
Conclusion and Future Outlook: Healthy Distance with AI
As clarified by the above analysis, current generative AI, particularly ChatGPT and Gemini, while possessing extremely sophisticated information processing capabilities, essentially remain “statistical reasoning machines” simulating “something like intelligence.”
6.1 Transition to System 2 Reasoning
The focus of future AI development is transitioning from intuitive word prediction (System 1) to more logical and thoughtful reasoning (System 2). Concepts like Google’s “Thinking Budget” attempt to reduce time-series errors and logical breakdowns by giving AI time to solve problems.
6.2 Required User Literacy
- Information Non-credibility: Adhere to the principle that AI answers should “not be taken entirely at face value,” especially ensuring human verification (cross-checking) for time, numbers, and factual relationships
- Setting Tool Boundaries: Position AI not as a friend or counselor, but as an “organizing, summarizing, and idea-support tool” for vast data
- Understanding Structural Vulnerabilities: Recognize that phenomena like mojibake and foreign language contamination are exposures of the system’s “internal limitations,” requiring prompt reconstruction or usage environment review when they occur
The above formula suggests a mathematical lower bound where errors in generative AI (Egen) are at least twice the misclassification rate of classification models (Emis)
Final Conclusion
AI is a powerful “exoskeleton” that extends our capabilities, but there is no “human” inside it. Recognizing that AI is “deliberately” created to that level is not an admission of technical limitations but the first step in building a healthy cooperative relationship between humans and machines. We should understand the computational scientific truth behind the “plausible words” AI presents and wisely utilize this “imperfect intelligence” as an excellent partner that stimulates curiosity.
