Beyond Feeling Better: Capability-Sustaining Emotional Dialogue as a Longitudinal Research Paradigm

The Problem With Current Emotional Support Systems

Emotional dialogue systems today optimize for one thing: making users feel better right now. The authors argue this misses a critical long-term risk. When systems prioritize immediate relief—validating distress, offering quick fixes—they may accidentally undermine users' ability to regulate emotions, make autonomous decisions, and maintain relationships independently. Over months or years of repeated use, this creates dependency rather than resilience.

The paper's core claim: emotional support dialogue should be redesigned around capability-sustaining goals. Instead of asking "Did this conversation reduce distress?", ask "Can this person cope better on their own afterward? Will they know when to stop using the system? Are they making self-endorsed choices?"

What the Research Reveals

The authors conducted two audits that expose how narrow current practice is. They reviewed 60 system-building papers in emotional dialogue using PRISMA (a structured review methodology) and found:

• 95% of papers chase relief—reducing negative emotion is the sole success metric.
• 0% evaluate capability outcomes—no studies track whether users' coping skills, decision-making autonomy, or social connection actually improved.
• Only 1 paper even mentions dependency, autonomy, or termination risk.

A second audit examined actual dialogue data from ESConv (a public emotional support conversation dataset). In supporter turns across 300 interactions:

• 43% contained capability-relevant functions (e.g., prompting reflection, building self-efficacy).
• 22% were generic suggestions with no tailoring.
• Only 4% used reappraisal (helping users reframe a situation), 6.7% built self-efficacy (confidence in their own ability), and a mere 0.3% set healthy boundaries.

The gap is stark: dialogues rarely equip users to help themselves.

The Proposed Framework

Capability-sustaining emotional dialogue (CSED) reorients the entire research pipeline—data collection, model training, system design, and evaluation—around longitudinal use. Instead of one-off conversations, the framework considers:

• Repeated use: Does the system reinforce helpful coping or create dependency?
• Non-use and transitions: When should users disengage? Are they ready?
• Termination: Does the system prepare users to succeed without it?

The authors propose six design commitments (e.g., foster autonomy, encourage social connection) mapped to four evaluation timescales (immediate, session-level, long-term behavioral, societal). They also release a protocol for extending their audit to analyze how trained models actually behave—a crucial step, since published papers don't reflect live system outputs.

Why This Matters

Emotional support systems are already deployed at scale; millions interact with them weekly. The current relief-focused design silently optimizes for engagement and user retention, not user flourishing. A system that keeps someone perpetually dependent on validation is "successful" by today's metrics but harmful by tomorrow's outcomes.

CSED makes the stakes explicit and testable. It demands longitudinal data, governance guardrails around termination, and honest evaluation of whether users are better equipped to live without the system. This shift—from feeling better to being better—redefines what emotional AI should aim for.