OpenAI has released a fresh safety assessment for its latest language models, introducing new testing protocols that examine how users might become emotionally dependent on AI systems.
The addendum focuses on both GPT-5.1 Instant and GPT-5.1 Thinking variants, establishing updated metrics for risk evaluation. The company's safety team conducted specialized evaluations targeting mental health considerations, recognizing a gap in how AI tools are assessed for potential psychological impact.
The new framework marks a shift in how the company approaches safety documentation. Rather than relying solely on traditional performance benchmarks, OpenAI is now incorporating measures that track whether users might rely too heavily on AI for emotional support or mental health guidance. This reflects broader industry conversation about the psychological effects of conversational AI.
The system card addendum details the company's testing results across these new categories, providing transparency about where the models perform well and where limitations exist. The Thinking variant, which uses extended reasoning, receives particular attention given its increased computational depth.
The move comes as AI systems become more integrated into daily life, with users increasingly turning to chatbots for advice on personal struggles. OpenAI's decision to formally document emotional reliance risks suggests the company views this as a legitimate safety concern rather than a fringe edge case.
The updated documentation is available to researchers and developers building on top of these models, allowing them to make informed decisions about deployment in sensitive contexts.
Author Emily Chen: "OpenAI is finally taking the mental health angle seriously, but publishing safety addendums only matters if developers actually read and act on them."
Comments