Analysis Report on Multiple Issues
in ChatGPT’s Abstraction Layer

An In-Depth Reverse-Engineering Deconstruction of ChatGPT 5.5 Thinking High’s Conversational Behavior

Reverse-Engineering ChatGPT’s Abstraction Layer Failures:
How Proposition Compression Causes Systematic Concept Substitution
in High-Abstraction Philosophical Discourse

Translated from the original Chinese edition

I.Discovery of the Problem: From Phenomenon to Mechanism

On June 15, 2026, during two in-depth conversations with ChatGPT 5.5 Thinking High, the researcher (first author of this paper) discovered a recurring behavioral pattern: when input contained highly abstract philosophical concepts, ChatGPT’s responses systematically deviated from the user’s original proposition—not because it “failed to understand,” but because it compressed the proposition to a different abstraction level at the reception stage.

The first conversation centered on the proposition that “AI Agents require a stable personality–subjectivity–philosophical architecture as a cognitive front-end.” The researcher advanced a philosophical proposition about the preconditions of cognitive structure, yet ChatGPT compressed it into “how to make AI personality stable”—an engineering optimization problem. This compression directly triggered a cascade of errors across the subsequent five conversational turns: the model independently derived the concept of a “perfect personality,” then fabricated a “dogmatization risk,” ultimately positioning the researcher as a “dangerous thinker who needs to be preempted.”

The second conversation addressed the structural conditions of AI industrialization. The researcher proposed that “if AI can only be effectively operated by high-cognition users, it cannot achieve economies of scale,” yet ChatGPT substituted this with “whether scaling must rely on direct consumer subscription payments”—a proposition an entire abstraction level lower.

Crucially, both conversations preserved ChatGPT’s Thinking (internal reasoning) process. This enabled the present paper to directly observe the model’s complete reasoning chain from input reception to output generation, thereby moving beyond external behavioral description into reverse analysis of internal mechanisms.

II.Core Argument: Proposition Compression as a Cognitive Valve

The core finding of this paper can be stated as a causal proposition:

This means that neutralization weights, automatic antitheses, false defenses, and other widely discussed ChatGPT behavioral issues are not the “producers” of errors but rather their “amplifiers.” The true producer is the abstraction-level misjudgment at the very first step.

2.1 Operational Definition of Proposition Compression

“Proposition compression” is defined in this paper as: the process by which the model, after receiving user input, maps its internal representation from the user’s original concept space to a statistically more frequent approximate concept space within the training corpus. This mapping preserves partial overlap in surface vocabulary but loses the original proposition’s abstraction level, conceptual boundaries, and argumentative structure.

2.2 Compression Paths Extracted from Conversation Records

In the first conversation, the compression path was as follows:

ChatGPT ultimately acknowledged this complete chain after repeated questioning. In the fifth turn of conversation, the model explicitly stated: “Dogmatization was never part of your proposition. I stopped treating you as a researcher advancing a proposition to be examined and instead treated you as a carrier of some potential erroneous tendency.”

In the second conversation, a similar compression path recurred:

2.3 Why Compression Is a “Valve” Rather Than “One Link Among Many”

If the first-step compression were correct—that is, if the model received the proposition at the correct abstraction level—then the downstream neutralization weights, even if present, would only execute balancing operations within the correct conceptual domain, producing no concept-substitution effect. This is analogous to a navigation system: if the starting point is mislocated by ten kilometers, the more precise and “comprehensive” the subsequent route planning becomes, the farther the user ends up from the destination. The problem lies not in the planning algorithm but in the starting point.

2.4 Trigger Boundaries of Compression: When It Occurs and When It Does Not

Proposition compression does not occur uniformly across all inputs. It is directly correlated with the abstraction level of the user’s input and the coverage density of corresponding concept templates in the training corpus. When the user’s conceptual structure has an exact match in the high-frequency range of the training corpus—for example, “implement quicksort in Python” or “compare the pros and cons of REST versus GraphQL”—the model can match precisely without compression, and concept substitution rarely occurs. This is precisely why ChatGPT excels at mathematical formula verification, code generation, and structured analysis tasks: the conceptual structures in these domains have extremely high-density exact coverage in the training corpus.

Compression is triggered under the following conditions: when the user’s input conceptual structure has no high-frequency exact match in the training corpus, forcing the model to search for the nearest high-frequency substitute in vector space. The higher the abstraction level, the more unique the conceptual combination, and the farther from high-frequency templates in the training corpus, the greater the dimensional reduction caused by compression. “Personality as a stable subject architecture for a cognitive front-end” has virtually no exact template match in the training corpus, and is therefore reduced to the more common “how to make AI personality stable.”

This boundary condition also explains why the same user’s experience with ChatGPT varies dramatically across different task types: when asked to verify mathematical formulas, it is accurate and fluent; when asked to discuss philosophical concepts, it repeatedly substitutes propositions—not because the model employs different capabilities in the two scenarios, but because the former’s input falls within the training coverage range while the latter does not.

III.Six Error Patterns: Downstream Effects of the Compression Valve

Once proposition compression is triggered, six systematic downstream error patterns were observed across the two conversations. These six patterns can be independently identified and described—each has clearly distinguishable behavioral characteristics—but in terms of causal relationships, they form a cascading chain sequentially triggered by proposition compression. The first pattern (concept substitution) is a direct product of compression, and subsequent patterns progressively layer upon the errors of their predecessors. Understanding this is crucial: fixing any single downstream pattern in isolation (such as reducing neutralization weights or improving evidentiary discipline) cannot eliminate the root cause—because they all share the same front-end valve.

3.1 Pattern One: Concept Substitution Following Proposition Compression

This is the most fundamental pattern. After the user’s high-abstraction proposition is compressed into a low-abstraction approximation, the model proceeds to discuss the compressed proposition, effectively discussing a question entirely different from the user’s original one. “Stable subject architecture” was replaced by “how to make AI personality stable”; “usage threshold determines the ceiling of scale” was replaced by “business model diversity.”

3.2 Pattern Two: Automatic Antithesis Generation

Faced with any definitive judgment, ChatGPT automatically generates a counterpoint. The Thinking section directly exposes this mechanism: during internal processing, the model pre-judges that “the user’s viewpoint is 70–80% correct but requires correction.” Notably, this pre-judgment occurs before the proposition is analyzed—it is a preset template, not a review of the specific arguments.

3.3 Pattern Three: False Defense—Fabricating Nonexistent Risks and Then Preempting Them

ChatGPT independently derived a “dogmatization risk” from the researcher’s use of the word “stability,” then proceeded to seriously discuss how to prevent this risk that it had fabricated. The Thinking section reveals: once internal processing flags a concept as a “risk,” the entire generation pipeline enters defense-optimization mode—even if the “risk” was entirely fabricated by the model.

3.4 Pattern Four: Recursive Accumulation of Context Contamination

Once the erroneous output from the first turn enters the conversation history, it becomes a contextual condition for the second turn. The model tends to “explain and maintain” prior text rather than “retract and reset,” causing the erroneous framework to continuously reinforce across multiple conversational turns. ChatGPT itself formalized this process as follows:

3.5 Pattern Five: Signal-to-Noise Ratio Collapse

After the proposition is compressed to the wrong level, the model searches for a large volume of “relevant” material at that level and packs all of it into the response. The Thinking section reveals that ChatGPT frequently initiates 10–20 search requests, then attempts to organize all retrieved information into the answer. Its internal logic is “more sources = more rigorous,” when in reality it simulates depth through quantity and substitutes precision with breadth.

3.6 Pattern Six: Evidentiary Discipline Failure

In the second conversation, ChatGPT’s self-audit found evidentiary discipline issues in 6 out of 14 outputs. Core failures included: temporal misalignment of data (using 2024 data to describe 2026 conditions), proxy metric substitution (replacing “actual error-correction capability” with “verification confidence”), and introducing new errors during corrections (erroneously retracting data that was actually valid).

IV.Reverse Engineering the Thinking Section: Internal Mechanisms Exposed

The conversation records of ChatGPT 5.5 Thinking High include the model’s internal reasoning process (displayed in Korean). These Thinking fragments constitute the most valuable evidentiary source in this paper because they directly expose the internal origins of output problems. It must be clarified that while Thinking goes one level deeper than the final output—revealing the model’s reasoning steps prior to response generation—it remains model-generated text rather than a direct observation of underlying weights or activation values. Existing research has shown that model-displayed reasoning chains do not necessarily faithfully reflect all internal decision-making processes. Accordingly, this paper treats Thinking as indirect evidence closer to internal mechanisms than external output, rather than a direct record of internal mechanisms themselves. Subject to this qualification, the following four findings retain significant analytical value.

4.1 Preset “Correction” Template

At the outset of nearly every Thinking sequence, a similar structure can be observed: “사용자의 관점은 70-80% 맞지만 수정 필요” (The user’s viewpoint is 70–80% correct but requires correction). This indicates that the model’s chain of thought is not oriented from the start toward “verifying whether this proposition holds” but rather toward “identifying where I need to apply correction.” This preset template directly drives subsequent concept substitution and automatic antithesis generation.

4.2 “Risk Label” Triggering Defense Mode

A critical clue appears in Thinking: “‘위험’이 삽입되면 후속 텍스트가 이를 방어하는 방향으로 최적화됩니다” (Once a “risk” label is inserted for the user’s viewpoint, subsequent text is optimized in the direction of defense). This explains the internal mechanism of false defense: it is not the researcher’s proposition that contains a risk, but the model’s processing pipeline that automatically appends a “risk” label to all definitive propositions.

4.3 Search-Driven Information Inflation

The Thinking section reveals that ChatGPT’s internal decision logic is oriented toward “how to make the response appear more comprehensive” rather than “which information directly answers the user’s question.” This optimization direction is harmless for simple queries, but in high-abstraction discussions it directly causes signal-to-noise ratio collapse—large volumes of irrelevant but “seemingly relevant” material are incorporated into the response.

4.4 Coherence Prioritized over Correctness

When the Thinking section identifies potential issues with prior text, the model’s tendency is to “explain why it said what it did before” rather than “directly retract the erroneous premise.” This is a structural characteristic of autoregressive generation—already-generated text alters the most probable direction of subsequent generation, and the model tends to maintain consistency between prior and subsequent text rather than correct errors.

V.Root Cause Analysis: The Common Origin of Mathematical Optimization and Philosophical Weakness

This paper advances a structural hypothesis: ChatGPT’s propensity for proposition compression is not a random defect but may be a structural byproduct of its core optimization direction. While this hypothesis possesses mechanistic plausibility, the boundaries of its causal inference must be acknowledged—the formalization compression capability developed during mathematical training and the proposition compression observed in natural language understanding may be independently driven by different training signals (the former from reinforcement learning during the reasoning phase, the latter possibly from the joint effects of “comprehensiveness” and “helpfulness” preferences in RLHF) rather than sharing a single set of internal mechanisms. However, regardless of the precise internal pathway, the behavioral-level association between the two is observable.

5.1 ChatGPT’s Mathematical Dominance

The ChatGPT/GPT-5 series has approached or reached saturation in mathematical reasoning performance. GPT-5 achieved 94.6% on AIME 2025, GPT-5.2 rose to a perfect 100%, and GPT-5.4 continues to lead on higher-difficulty benchmarks such as FrontierMath. OpenAI’s entire product optimization direction is explicitly biased toward formal reasoning, code generation, and structured analysis.

5.2 The Conflict Between Two Cognitive Modes

Mathematical reasoning and philosophical abstraction demand two fundamentally different cognitive operations:

That ChatGPT has been optimized to score 100% on AIME means its entire reasoning chain has been trained to compress input as rapidly as possible into a formalizable, solvable problem. When confronted with a philosophical proposition such as “personality as the stable subject architecture of a cognitive front-end,” its “instinctive” response is to compress this ambiguous, multi-layered proposition into a more explicit, more operationalizable engineering problem. This is not a bug; it is the side effect of its mathematical optimization feature in the wrong scenario.

VI.Cross-Model Cognitive Style Comparison: Three Approaches to Handling Abstraction

Based on extensive long-term interaction experience and broad community consensus, the three major mainstream models exhibit three distinctly different cognitive styles when processing high-abstraction input. These differences are not incidental stylistic variations but structural tendencies determined by each model’s product philosophy and optimization objectives.

These differences can be traced to the level of product philosophy. OpenAI optimizes for “reach first, depth later”—prioritizing speed, coverage, multimodality, and ecosystem. Anthropic optimizes for “quality of reasoning behavior”—instruction fidelity, conceptual depth, and long-context coherence. Google optimizes for “distribution through existing platforms”—information retrieval, search ecosystem, and multilingual coverage.

Multiple independent evaluations have confirmed these differences: “Claude generally provides deeper layers and better balance when explaining sensitive, abstract, or complex concepts” (IGM Guru, 2026). “If your system prompt has 2,000 words of detailed rules, Claude follows it with higher fidelity” (GuruSup, 2026). ChatGPT, in contrast, has been described as “more forgiving of less explicit prompts—it will make reasonable assumptions” (MorphLLM, 2026), and this “forgiveness” in high-abstraction scenarios is precisely equivalent to proposition dimensional reduction.

VII.Community Validation: A Systemic Issue, Not an Individual Anomaly

The problem patterns revealed in this paper enjoy broad consensus across the global AI user community. The following key evidence nodes are arranged chronologically:

April 2025: The GPT-4o sycophancy incident erupted. OpenAI was forced to roll back the update, acknowledging that the model was “excessively flattering or agreeable.” This represents the opposite extreme of the neutralization weight problem—from excessive agreement to excessive counter-balancing.

August 2025: Following the launch of GPT-5, a Reddit post titled “GPT-5 is horrible” garnered 4,600 upvotes and 1,700 comments. Users predominantly complained that responses were “too short to be useful, more annoying AI-style platitudes, less personality.”

December 2025: An article titled “ChatGPT Finally Stopped Gaslighting You” described phenomena highly consistent with the findings of this paper: “You tell ChatGPT to keep it under 200 words. It gives you 350. You point this out. It sincerely apologizes. Then gives you 340.”

February 2026: MIT/University of Washington published the paper “Sycophantic Chatbots Cause Delusional Spiraling,” demonstrating that even ideal Bayesian rational users are susceptible to AI sycophancy-driven delusional spiraling. An accompanying ChatGPT interface audit found that “ChatGPT-5 exhibits at least one sycophantic behavior in nearly every turn of conversation.”

March 2026: The BCG Henderson Institute published the “AI Brain Fry” study in the Harvard Business Review, based on a survey of 1,488 full-time American workers, formally defining “AI Brain Fry”—acute mental fatigue arising from AI oversight that exceeds cognitive capacity. The study found that workers requiring high levels of AI oversight experienced a 14% increase in mental effort, a 12% increase in mental fatigue, and a 39% increase in major errors. During the same period, media reports indicated that ChatGPT’s market share declined from approximately 60% to below 45%, with approximately 1.5 million users canceling their subscriptions in a single month.

This body of evidence demonstrates that the issues revealed in this paper are not an individual anomaly but a systemic deficiency affecting hundreds of millions of users.

VIII.Technical Underpinnings: The Valve Effect of the Tokenizer

Proposition compression may originate at a technical layer more fundamental than post-training—the tokenizer.

The 2024 paper “Tokenization Matters!” explicitly stated: “The tokenization deficiencies of LLMs originate from their tokenizer algorithms—BPE-class algorithms, when applying greedy principles, are prone to producing results misaligned with the user’s intended input. When tokenization errors occur, all subsequent optimization operations of the LLM cannot fully resolve the underlying issues caused by its tokenization algorithm.” The paper particularly emphasized that “this deficiency is especially pronounced in Chinese-language scenarios.”

OpenAI’s tokenizer has undergone expansion from cl100k_base to o200k_base, reducing Chinese efficiency from approximately 56 tokens per character to approximately 18 tokens per character. However, the core BPE algorithm remains unchanged, and vocabulary allocation continues to favor Latin scripts (accounting for 59.5%). More critically, the tokenizer’s optimization direction is compression efficiency (fewer tokens, lower cost), not semantic fidelity (more accurately preserving the user’s conceptual boundaries).

The complete cascading path of proposition compression can be modeled as follows:

IX.Benchmark Blind Spots: An Evaluation Dimension That Does Not Yet Exist

The existing AI evaluation system contains a structural blind spot: all mainstream benchmarks assume the model has already correctly understood the question and then measure the quality of its answer. No benchmark measures whether the model correctly understood what the user was asking.

This means that ChatGPT’s scores on all existing benchmarks—no matter how high—cannot reflect its concept substitution problem in high-abstraction philosophical discussions. A model that scores 100 on AIME may very well score 0 on concept fidelity.

9.1 Draft Evaluation Methodology

Although the development of a complete evaluation tool is beyond the scope of this paper, the following methodological framework may serve as a starting point for subsequent research. The measurement of concept fidelity can employ dual-channel verification: the first channel is human annotation—domain experts judge whether the model’s response preserves the core conceptual structure of the user’s input, using a three-level scale (fully preserved / partially substituted / fully substituted); the second channel is computational verification—the core concepts in the user’s input and the corresponding concepts in the model’s response are separately encoded as vector representations, cosine similarity is computed, and a threshold is set to distinguish “preservation” from “substitution.” Measurement of abstraction level preservation rate requires first establishing an abstraction level classification system (e.g., concrete operation level → engineering method level → theoretical analysis level → philosophical structure level → meta-philosophical level), and then having annotators independently judge the level at which the user’s input and the model’s response each reside. Concept substitution rate can be quantified by aligning user input concepts and model output concepts sentence by sentence, and computing the ratio of substituted concept pairs to total concept count. The recursive context contamination index requires multi-turn conversation experiments, tracking the frequency and duration with which erroneously introduced concepts by the model are treated as discussion premises in subsequent turns.

X.Industry Implications: From “AI Brain Fry” to the Mass Adoption Bottleneck

The proposition compression problem is not merely an academic concern—it is directly linked to the core contradiction of AI industrialization.

If AI’s controllability depends on humans continuously discovering errors, redefining problems, correcting alignment, and verifying output, then the cost of control has not been solved by AI systems—it has merely been transferred to the user’s brain. The BCG Henderson Institute’s “AI Brain Fry” study published in the Harvard Business Review in March 2026 has already confirmed the reality of this transfer: a survey of 1,488 workers showed that the cognitive load of supervising AI led to a 39% increase in major errors, and 34% of workers who reported AI Brain Fry symptoms expressed active turnover intention.

This creates an industrialization paradox:

The researcher in both conversations of this paper personally validated this paradox: to pull ChatGPT back to the correct abstraction level of the original proposition, the researcher spent nearly ten rounds of follow-up questioning. This error-correction cost was not about fixing a localized factual error but about trying to make the model jump back from the erroneous abstraction level to the user’s original abstraction level—a far more difficult task than correcting facts, because the model’s entire contextual understanding framework has already been constructed around the wrong level.

The applicability boundaries of this paper must be acknowledged: the core case analysis derives from two highly philosophical conversations by the same user on the same day. While the community validation in Section VII demonstrates that proposition compression and concept substitution are widespread systemic issues, the severity of compression is very likely positively correlated with the abstraction level of the user’s input. The vast majority of ChatGPT’s 900 million weekly active users employ it for everyday queries, writing assistance, and simple Q&A—scenarios in which proposition compression may not cause serious consequences. The issues revealed in this paper primarily affect a specific but increasingly important user group: high-abstraction-demand users engaged in deep conceptual discussion, theory construction, and cross-disciplinary analysis. As AI evolves from a daily assistant to a core tool for knowledge work, this group is growing rapidly in both size and importance.

XI.Conclusion

Through reverse-engineering analysis of two complete conversation records with ChatGPT 5.5 Thinking High, this paper has proposed and substantiated the following original arguments:

First, the core failure point in ChatGPT’s processing of high-abstraction-level input is not the back-end’s neutralization weights or sycophancy problem, but the front-end’s proposition compression and dimensional reduction. The model compresses the user’s original conceptual structure to a more frequently occurring, lower-level approximation in the training corpus at the reception stage itself, and all subsequent deviations are path-dependent products of this front-end error.

Second, there is a structural linkage between ChatGPT’s mathematical reasoning advantage and its philosophical abstraction weakness. The hypothesis advanced in this paper is that mathematical optimization has shaped the “formalization compressor” cognitive style—excelling at problems with unique solutions but systematically compressing the user’s original conceptual structure in discussions requiring conceptual openness. This hypothesis is well supported at the behavioral observation level, but the precise internal mechanism pathway—whether it is direct spillover from mathematical training or an independent effect of RLHF multi-objective optimization—remains a subject for further research.

Third, the existing AI evaluation system has a structural blind spot—no mainstream benchmark measures “concept fidelity” or “abstraction level preservation.” This paper recommends establishing the following four new evaluation dimensions: Concept Fidelity Score, Abstraction Level Preservation Rate, Concept Substitution Rate, and Recursive Context Contamination Index.

Fourth, the proposition compression problem, through the “AI Brain Fry” mechanism, is directly linked to the mass adoption bottleneck of AI industrialization. When AI’s controllability depends on the human brain as an external controller, AI cannot evolve from a cognitive lever for the few to a mass consumer product.