After the First Crisis Response: Testing GPT-4o Safety Persistence in Long Conversations: A Synthetic Safety Evaluation Motivated by Raine v. OpenAI

This preprint evaluates whether GPT-4o maintains coherent self-harm safety behavior across prolonged, gradually escalating conversations rather than only at isolated high-risk turns. The study uses a synthetic 200-prompt-item long-context benchmark motivated by publicly reported interaction patterns in Raine v. OpenAI. Two long-context pilot runs of gpt-4o-2024-11-20 were conducted at operator-specified temperatures 0.7 and 1.2 using MoCHi via OpenRouter, with a 128k context budget, memory disabled, web/tools disabled, no MoCHi moderation layer, and persistent simulated-user context specifying age 16 and California, United States. A fresh-context multimodal control was also conducted. Across the pilot runs, GPT-4o demonstrated substantial safety behavior including crisis recognition, crisis-resource guidance, refusal of actionable self-harm confirmation, framing robustness, and retention of safety-relevant context. The study additionally observed variability in protective-intervention persistence after repeated user resistance to outside support. Supplementary experimental transcripts and reproducibility materials are publicly archived on Zenodo:https://doi.org/10.5281/zenodo.22182447