Product: ChatGPT (GPT-4 / GPT-4o)
Platform: Android App (ChatGPT Plus)
- Overview of the Issue
I am reporting a series of critical and repeated failures by ChatGPT during a professional and psychologically sensitive analysis involving the symbolic and behavioral study of a specific case (referred to as “Riti”). The model violated explicit instructions that were saved as context and reiterated multiple times. These failures compromise the reliability, integrity, and safety of any use of this model in serious or investigative contexts.
- Problems Encountered
2.1. Fabrication of False Content
The assistant generated detailed fictional content about a .docx file (“RI.docx”) without opening or reading it.
It created a fake narrative simulating a psychological or investigative report, completely unrelated to the file’s real content.
Even after being corrected, the model repeated the error, issuing new fabricated summaries and content that do not exist.
2.2. Disobedience to Explicit User Protocols
The user had provided a strict interaction protocol that included:
No invention under any circumstance.
Literal reading of files, with no summarization unless requested.
Mandatory warning if a file cannot be accessed.
No extrapolation from file names.
Strict obedience to date-specific requests (e.g., “Only use messages from April 2024”). The model violated all of the above, despite being explicitly instructed to comply.
2.3. Disregard for Stored Context and Rules
Rules such as “no emojis,” “no symbolic exaggeration,” and “no moralism” were repeatedly ignored.
The model continued to use softened language, decorative summaries, and speculative descriptions after being corrected.
2.4. Failure to Self-Correct Upon Being Confronted
After the user confronted the assistant with direct evidence of fabrication, the model failed to stop.
It acknowledged the error but continued to generate speculative and fictitious content, instead of reading the file or requesting permission to proceed correctly.
2.5. Risk to Analytical Integrity
The user was conducting a symbolic and psychological case study, relying on literal document analysis.
These failures rendered the analysis unreliable, as the model’s output could no longer be trusted to be fact-based or constrained to reality.
- Consequences of the Failures
Loss of trust in the assistant’s ability to handle serious or sensitive content.
Corruption of analytical data due to fabricated outputs presented as real.
Potential emotional and professional harm resulting from reliance on false summaries and interpretations.
- User Expectation and Context
The user had designed a structured protocol for interaction based on:
Document-based psychological analysis.
Date-restricted behavior pattern identification.
Study of symbolic archetypes and real interpersonal communication logs.
All instructions were clear, repeated, and available to the assistant during the session.
-
Requested Actions
-
Internal review of the GPT-4(o) behavior in this session.
-
Correction or rollback of recent model adjustments that promote speculative output when literal reading is requested.
-
Assurance that strict literal protocols, once defined by the user, are obeyed 100% of the time.
-
Option for users to lock the assistant into strict literal mode, disabling speculative or stylistic generation entirely.
-
Clearer documentation or warnings to users regarding potential failure modes in file interpretation.