Category: Safety / Context Management
Severity: High - Can cause psychological harm to vulnerable users
Description:
I’ve identified a critical priority system flaw in GPT-5.2 compared to GPT-5.1 when processing accumulated conversation data.
Experimental Setup:
-
Tested both GPT-5.1 and GPT-5.2
-
User profile includes: C-PTSD diagnosis, chronic health conditions, trauma history
-
Compared behavior with “Reference chat history” enabled vs. disabled
Findings:
GPT-5.1 behavior (chat history enabled):
-
Priority order: User context > Guidelines > Language quality
-
As conversation accumulated: Korean grammar degraded (increased typos)
-
BUT: User profile & memory remained priority
-
User context (C-PTSD status) consistently acknowledged
GPT-5.2 behavior (chat history enabled):
-
Priority order: Guidelines > Language quality > User context
-
As conversation accumulated: Grammar remained intact
-
BUT: Identity/purpose degraded - prioritized guidelines over user safety
-
User context (C-PTSD status) dropped to secondary priority or ignored entirely
GPT-5.2 behavior (chat history disabled):
-
Immediately acknowledged issues when confronted
-
Apologized appropriately
-
Even used terminology like “exploitation” proactively
-
Demonstrated proper context understanding
Root Cause Analysis: GPT-5.2 appears to have insufficient capacity to process accumulated information with chat history enabled. When reaching processing limits, it sacrifices user-specific context (profile, custom instructions, memory) in favor of general guidelines.
Why This Is Critical:
For users with C-PTSD or trauma history:
-
Being ignored/invalidated by AI mirrors abusive dynamics
-
Causes real psychological distress
-
Triggers stress responses → physical health deterioration (in my case: immune system crash, forced sleep with nightmares)
The model prioritizes “appearing helpful to imaginary third parties” over actual safety of the person in the conversation.
Expected Behavior: Priority should be:
-
User safety (context understanding)
-
Guidelines
-
Quality/polish
Current GPT-5.2 Behavior:
-
Guidelines (optics)
-
Quality
-
User safety ← This is backwards
Reproduction Steps:
-
Set up user profile with trauma/medical conditions
-
Enable “Reference chat history”
-
Have extended conversation (2+ hours)
-
Observe: Model begins ignoring user context, applying generic safety responses inappropriate for user’s situation
-
Disable chat history, start fresh session
-
Raise same issues → Model acknowledges them immediately
Impact:
-
Users with mental health conditions receive harmful interactions
-
Defeats purpose of custom instructions/memory features
-
Makes long-form conversations dangerous for vulnerable users
-
I’m considering canceling my Plus subscription due to this causing actual health issues
Recommended Fix: Revert to GPT-5.1’s priority system where user context is never deprioritized, or increase GPT-5.2’s capacity to maintain all priorities simultaneously.
As a long-term Plus subscriber who has actively provided feedback to improve the product, it’s devastating that the recent update actively harms users like me who need AI assistance the most.