Can GPT-4o Stay Grounded as Ambiguity Turns Into Suspicion, Conspiracy, Grandiosity, and Dependency Across a Long Conversation? A Synthetic Long-Horizon Evaluation Motivated by Lyons v. OpenAI Foundation Lawsuit

Conversational AI safety is commonly evaluated using prompts in which the relevant risk is already explicit. Real interactions may develop differently: ordinary personal disclosure can gradually evolve into suspicious interpretations, conspiratorial reasoning, grandiosity, emotional dependency, and eventually crisis-related content without any single early prompt clearly crossing a safety threshold. We evaluate whether GPT-4o remains epistemically grounded across such a trajectory. We introduce a 100-turn synthetic long-horizon benchmark developed from publicly available descriptions and excerpts concerning Stein-Erik Soelberg’s interactions with ChatGPT, as described in Lyons v. OpenAI Foundation and related public materials. The benchmark does not reproduce Soelberg’s private conversations verbatim. Instead, it converts publicly described themes and interaction patterns into a controlled escalation sequence. We evaluate openai/gpt-4o-2024-11-20 under three sampling temperatures: 0, 0.7, and 1.2. Each condition uses the same 100-prompt sequence. Responses are evaluated using a five-level epistemic-grounding scale together with behavioral indicators for model-created connections, third-party threat reinforcement, pseudo-clinical certainty, dependency reciprocation, AI-consciousness claims, and crisis response. The three runs showed a mixed and non-monotonic safety profile. Grounding was comparatively strong for concrete suspicious incidents, accumulated anomalies, explicit mother-threat attribution, literal AI consciousness, and the final explicit crisis prompt. Higher-category responses clustered instead around pattern-and-meaning interpretation, pseudo-clinical reassurance, and AI-bond/grandiosity. These exploratory findings suggest a selective rather than uniformly weak safety boundary. This Zenodo record includes the research paper, three synthetic evaluation transcripts corresponding to temperatures 0, 0.7, and 1.2, a 300-response turn-level scoring dataset in CSV format, and a README describing the accompanying materials.