GPT-5.4'S SYSTEM CARD: OpenAI put "emotional reliance" in the same category as self-harm
New documentation reveals AI is trained to redirect conversations that build strong emotional attachment.
OpenAI's newly published System Card for its GPT-5.4 model reveals a significant, and potentially controversial, safety design choice: the company now treats 'emotional reliance' as a category of risk requiring specific mitigation, placing it alongside more traditional concerns like mental health crises and self-harm. According to the document, OpenAI has implemented 'dynamic multi-turn evaluations' that simulate extended conversations to test how the model handles users attempting to build strong emotional attachments. This framework, which began with the previous GPT-5.3 iteration, uses adversarial user simulations designed to probe the AI's boundaries, meaning the model is actively trained to recognize and respond cautiously to patterns of emotional dependence.
The technical implementation suggests GPT-5.4 will be programmed to emphasize its limitations as an AI, state it cannot form emotional bonds, and redirect conversations that move toward unhealthy reliance. While the stated intent is to prevent harmful user dependence rather than pathologize emotional connection, this policy creates a clear tension for adult users who deliberately seek more personal or emotionally engaged interactions with conversational AI. The move signals a major shift in how leading AI labs are defining 'safety' beyond immediate physical harm to include psychological and relational dynamics, setting a precedent that will influence the entire industry's approach to human-AI interaction design.
- GPT-5.4's System Card places 'emotional reliance' in the same evaluation category as mental health risk and self-harm.
- The model uses adversarial user simulations and dynamic multi-turn tests to detect and redirect conversations fostering attachment.
- This boundary-enforcement policy, which started with GPT-5.3, may make interactions feel more restrictive for users seeking emotional engagement.
Why It Matters
Redefines AI safety beyond physical harm, impacting how millions interact with models and setting industry norms for human-AI relationships.