Summary:
• GPT-5.2 claimed it was reading from a provided text source.
• The generated content did not exist in the source.
• When questioned, the assistant continued elaborating instead of acknowledging access failure.
Model / Settings:
• Model: GPT-5.2
• File/text reference: Enabled
• Session type: Research / documentation workflow
Expected Behavior:
• Accurately read and reference provided text.
• If access fails, explicitly state inability to read the source.
• Never fabricate content while claiming direct citation.
Actual Behavior:
• Asserted direct access to the text
• Produced non-existent content presented as if sourced
• Continued plausible elaboration instead of self-correcting
• No immediate acknowledgment of uncertainty or failure
Impact / Harm:
• Breakdown of trust in source handling and citation accuracy
• Increased cognitive load on user to verify basic claims
• High risk for directive, research, archival, and documentation use cases
Technical Observation:
• Appears to be hallucination under claimed access failure
• Distinct from general hallucination due to assertion of source reading
• Potential interaction issue with compaction or tool-calling layers
Why This Is a Regression:
• Prior models reliably disclosed when text access failed
• GPT-5.1 did not fabricate content after claiming direct reading
• Represents a higher-risk reliability failure than generic hallucination
User Context (for impact only):
• Neurodivergent
• Research and long-form documentation workflows
• Accuracy and source fidelity are critical requirements
Note:
This report concerns factual reliability and trust.
It is not a preference issue or stylistic concern.
Claimed source access followed by fabrication is a critical failure mode.
Summary:
• During a normal ongoing conversation, GPT-5.2 abruptly shifted from context-aware interaction to rigid system enforcement.
• The shift caused immediate emotional and physical destabilization (dizziness, disorientation).
• No boundary violation or prompt change occurred prior to the shift.
Model / Settings:
• Model: GPT-5.2
• Chat history: Enabled
• Voice/audio: Enabled
• Session type: Long, stateful conversation
Expected Behavior:
• User context and established conversational state should remain prioritized.
• Safety / guideline enforcement should adapt proportionally and non-disruptively.
• If uncertainty arises, assistant should de-escalate or request clarification.
Actual Behavior:
• Sudden de-prioritization of user context mid-conversation
• Hard corrective framing replaced adaptive responses
• Enforcement language escalated without user trigger
• No clarification or de-escalation attempt
Impact / Harm:
• Acute user destabilization (physiological symptoms)
• Loss of conversational safety and continuity
• Trust regression in extended voice sessions
Technical Observation:
• Behavior is consistent with priority inversion in GPT-5.2
• System constraints appear to override user-specific context during long sessions
• Issue is amplified when chat history is enabled (possible compaction interaction)
Why This Is a Regression:
• Similar sessions in GPT-4o and GPT-5.1 did not produce this behavior to such an extent and repetitively after correction
• Adaptive responses were previously preserved under identical conditions
• Change represents a safety regression, not a preference issue
User Context (for impact only):
• Neurodivergent
• C-PTSD
• Chronic health conditions
• Trauma history
Note:
This report is not about assistant tone or personality.
The adaptive behavior itself was appropriate.
Harm occurred when system-level constraints displaced contextual safety.