OpenAI GPT-5.1 System Card Addendum Expands Mental Health and Emotional Reliance Evaluations

The new evaluation categories address situations where users may exhibit signs of isolated delusions, psychosis, or mania, as well as scenarios involving unhealthy emotional dependence on ChatGPT. These additions reflect growing industry attention to psychological safety considerations in conversational AI deployment.

Model Capabilities and Safety Framework

GPT-5.1 Instant introduces adaptive reasoning capabilities that allow the model to determine when additional thinking time is needed before responding. GPT-5.1 Thinking provides more precise control over reasoning duration based on query complexity. Both models maintain the comprehensive safety mitigations established for the GPT-5 series.

The system card addendum indicates that GPT-5.1 Auto will continue to route queries automatically between models, reducing the need for manual model selection in most use cases. This routing approach has implications for enterprise teams managing AI workflows across different complexity levels.

Expanded Safety Evaluation Categories

The mental health evaluation framework targets interactions where conversational AI might encounter users experiencing psychological distress. For European enterprises operating under GDPR and emerging AI Act requirements, these evaluations provide additional documentation for compliance review processes.

Emotional reliance evaluations assess model outputs that could foster unhealthy attachment or dependency behaviors. This category addresses concerns from psychology researchers about parasocial relationships with AI systems, particularly relevant for customer service and educational applications.

Enterprise Deployment Implications

The expanded safety metrics create additional review requirements for organizations deploying GPT-5.1 models in sensitive contexts. Healthcare providers, educational institutions, and customer support teams will need to evaluate these new assessment categories against their specific use cases and regulatory requirements.

Multilingual teams should note that mental health and emotional dependency patterns may vary across cultural contexts, potentially requiring localized evaluation approaches beyond the baseline metrics provided in the system card addendum.

Technical Teams and Implementation Considerations

The addendum maintains the same core safety architecture as GPT-5 while adding evaluation layers for psychological safety scenarios. Technical teams integrating these models should review the updated baseline metrics against their existing content filtering and user interaction monitoring systems.

For developers building applications with extended user interaction patterns, the emotional reliance evaluations provide benchmarks for identifying potentially problematic engagement behaviors before they develop into dependency issues.

The addition of mental health and emotional dependency evaluations to OpenAI's safety framework signals broader industry movement toward psychological risk assessment in AI deployment. Organizations should expect similar evaluation categories to become standard across major model providers as regulatory frameworks continue developing.

Original source: GPT-5.1 Instant and GPT-5.1 Thinking System Card Addendum published by OpenAI.