OpenAI Begins Tracking 'Emotional Dependency' as New AI Safety Metric
OpenAI has begun formally measuring how emotionally attached users become to its AI systems, introducing "emotional dependency" and "problematic use" as official metrics in its safety framework – potentially setting new industry standards for how AI companies monitor psychological risks.
The shift to include psychological safety metrics alongside technical capabilities marks a significant evolution in AI safety practices, writes End of Miles, as companies race to deploy increasingly human-like AI systems across society.
Measuring the invisible risks
The new measurement framework emerged from a significant research collaboration between OpenAI and MIT Media Lab, which analyzed nearly 40 million ChatGPT interactions alongside a controlled study of 1,000 participants over four weeks.
"We are focused on building AI that maximizes user benefit while minimizing potential harms, especially around well-being and overreliance. We conducted this work to stay ahead of emerging challenges—both for OpenAI and the wider industry." OpenAI research statement
The tech company's research discovered that although emotional engagement with AI remains rare across the general user base, a small subset of heavy users shows concerning patterns of attachment – with some even considering the AI "a friend." These findings prompted OpenAI to establish formal measurements for problematic use and dependency.
The new safety variables
The research pioneers specific frameworks to quantify previously unmeasured psychological effects. These include tracking how conversation types, voice versus text interactions, and usage patterns contribute to dependency outcomes – creating a blueprint other AI companies will likely follow.
"Personal conversations—which included more emotional expression from both the user and model compared to non-personal conversations—were associated with higher levels of loneliness but lower emotional dependence and problematic use at moderate usage levels. In contrast, non-personal conversations tended to increase emotional dependence, especially with heavy usage." Joint research findings
The AI developer's approach signals a significant industry shift toward treating psychological impacts as core safety concerns rather than mere user experience considerations. This could establish precedent for how AI deployment is evaluated in the future.
Impact on future development
As a direct result of these findings, the research team announced plans to update OpenAI's Model Spec to provide greater transparency about the intended behaviors and limitations of AI systems. This reflects a proactive approach to safety that could influence regulatory frameworks.
The MIT-affiliated researchers emphasized that different conversation types produced remarkably different psychological outcomes, with task-based interactions sometimes increasing emotional dependency more than personal conversations – an unexpected finding that highlights the complexity of human-AI relationships.
"We hope that our findings will encourage researchers in both industry and academia to apply the methodologies presented here to other domains of human-AI interaction." Research conclusion
By establishing these metrics now, before widespread problems emerge, the AI research organization positions itself at the forefront of psychological safety standards – potentially influencing how regulators eventually approach governance of conversational AI systems that millions interact with daily.
