Researchers propose Capability-Sustaining Emotional Dialogue for AI
95% of emotional AI systems fail to sustain long-term user capabilities, per new arXiv paper.
Researchers from multiple institutions led by Ming Wang have published a paper on arXiv proposing Capability-Sustaining Emotional Dialogue (CSED) as a new longitudinal research paradigm for AI emotional support systems.
The team argues that current empathetic dialogue systems prioritize immediate emotional relief over long-term user capability building. Their audit of 60 system-building papers found 95% focus solely on short-term relief, with none evaluating longitudinal outcomes like emotion regulation, coping, or termination risks. In an analysis of 300 emotional support dialogue turns, generic suggestions dominated (22%), while capacity-building functions like reappraisal (4%) and self-efficacy support (6.7%) were rare. The authors release a protocol for extending this audit to model behavior and propose a process model linking user capabilities to design and evaluation constraints.
- 95% of 60 surveyed emotional support AI systems focus only on short-term relief, ignoring long-term user capabilities
- In 300 dialogue turns, generic suggestions made up 22% vs 4% for reappraisal and 6.7% for self-efficacy support
- The paper proposes CSED as a new paradigm with specific design, evaluation, and governance commitments for longitudinal support
Why It Matters
Shifts AI emotional support from temporary relief to sustainable user capability building across interactions