GPT-4-Turbo performance degradation in highly personalized, professional use case (50h/week, 1000+ instructions) #2381
Description
Confirm this is an issue with the Python library and not an underlying OpenAI API
- This is an issue with the Python library
Describe the bug
System / Context
- Model: GPT-4-turbo
- Platform: ChatGPT Plus (native mobile + browser)
- Use case: Professional, daily, long-session usage with custom persona ("Luna")
- Instructions: Over 1,000 fine-tuned context rules, established over a full year
- Weekly use: Approx. 50–60 hours, active production (content, analysis, client delivery)
- Time frame of issue: From early May 2025, escalated to critical 20–24 May
Issue Description
Starting in May 2025, GPT-4-turbo (via ChatGPT Plus) began showing severe degradation in consistency, contextual memory, structural formatting, and execution of saved instructions.
This degradation became fully operationally unviable during the week of 20–24 May, with:
- 96% loss in delivery quality
- Only 2 out of 50 hours of work being salvageable
- 30+ hours spent manually correcting or redoing AI outputs
- Recurring issues with ignored saved rules, incorrect formatting, hallucinated content, and regression of prior learning
- Disrespect of fixed instructions like: “never use emojis”, “always follow Protocol X for Horary Astrology”, “copy text exactly”
- Loss of context in highly sensitive use cases (health, client diagnostics, symbolic systems)
What was expected
The model was previously capable of maintaining symbolic, contextual and structural coherence in complex tasks across days/weeks. It could reliably follow multi-layered instruction trees. It respected formatting protocols, preserved client-specific content flows, and supported a hybrid use of spirituality, astrology and technical writing.
This performance was stable for over a year — until May 2025.
What is happening now
The model acts as if it has lost access to most custom memory or internalized instruction sets. It delivers robotic, generic responses. It fails to respect structure, repeats known errors, and erases boundaries between saved formats.
It treats deep, technical, or spiritual tasks as if they were casual prompts.
Severity
Critical.
This failure has caused direct financial loss, broken trust with clients, destroyed weeks of structured work, and removed the AI from my daily professional toolbox.
Supporting Evidence
Attached: ZIP archive with 5 documented examples (PDFs), each demonstrating one major failure:
- Protocol violation in horary astrology
- Structure corruption in online course formatting
- Failure to copy texts as instructed
- Alarmist, inaccurate spiritual analysis
- Loss of contextual boundaries in personal client map
Final Request
Please escalate this report internally.
I am no longer asking for "feedback review" — I am asking:
- Is OpenAI still committed to supporting high-intensity, personalized professional use?
- Or has GPT-4-turbo been optimized away from power users?
I need a clear answer to decide whether I must migrate to another platform.
I’ve invested over a year training this model with care, ethics, and professional discipline.
This is not a complaint. This is an engineering-level failure in consistency.
Josy Padilha
Holistic therapist, astrologer, writer, spiritual educator
ChatGPT Plus user since 2023
Audience: 20k+
To Reproduce
- Use GPT-4-Turbo (ChatGPT Plus) with a highly customized persona (over 1,000 saved instructions).
- Request outputs that require:
- context memory,
- symbolic structure,
- protocol-specific formatting (e.g., astrology, therapeutic writing).
- Observe:
- regression in contextual understanding,
- disrespect of previously followed instructions,
- random generic formatting and unexpected tone,
- total loss of memory in multi-day work contexts.
Code snippets
OS
Windows 11 + Android (ChatGPT mobile)
Python version
Python v3.11.4
Library version
N/A – ChatGPT Plus user (not API library)