AI Calms Panic Attacks Better Than Humans

Panic attacks strike without warning, leaving millions struggling to breathe and convinced they're dying. Now, researchers have developed an artificial intelligence system that provides immediate, effective counseling during these terrifying episodes—outperforming both human therapists and other AI models in crisis situations.

The key breakthrough is PACER, an AI counseling agent specifically trained to deliver Psychological First Aid during acute panic attacks. Unlike general-purpose counseling AI or traditional therapy bots, PACER focuses exclusively on rapid stabilization through clear, directive guidance. The system was trained on PACE, the first dataset specifically designed for panic attack scenarios, containing over 8,000 synthetic counseling dialogues grounded in real-world panic experiences.

Researchers built PACER using a novel training approach that combines supervised learning with preference alignment. The AI first learned from high-quality counseling examples, then refined its responses through simulated interactions with panicked users. This dual training method taught PACER to provide empathetic yet directive support—guiding users to safer locations, leading breathing exercises, and offering reassurance without overwhelming them with complex questions.

Experimental results demonstrate PACER's superiority in crisis management. When evaluated using the VAL framework—a new multi-dimensional assessment tool covering both general counseling skills and panic-specific interventions—PACER outperformed all baseline models. It achieved higher scores in clarity (4.96/5), directive support (4.735/5), and emotional stabilization (4.925/5) compared to general counseling AI, CBT-based systems, and even the powerful GPT-4o model. Most impressively, PACER stabilized users in just 7.17 turns on average, significantly faster than other approaches.

The real-world impact is measurable through emotional changes. Using the Positive and Negative Affect Schedule, researchers found PACER produced the largest reductions in core panic emotions—distress and fear decreased sharply, while self-related positive emotions like pride and determination increased. Human evaluations confirmed these findings, with both clinical experts and individuals with lived panic experience consistently preferring PACER over alternatives in head-to-head comparisons.

This technology matters because panic attacks affect approximately 11% of adults annually, yet immediate professional help is often unavailable due to cost, timing, or accessibility constraints. PACER offers 24/7 support that's specifically tailored to high-distress situations where traditional therapeutic approaches may fail. The system's action-oriented guidance—moving users to safer spaces, leading breathing exercises, providing immediate reassurance—addresses the urgent needs of someone in crisis.

However, limitations remain. The research acknowledges that PACER shows slightly lower empathy scores than GPT-4o, suggesting a trade-off between immediate support and emotional resonance. Additionally, while the system was extensively tested through simulations and expert evaluations, it hasn't been deployed with real users experiencing actual panic attacks. The researchers also implemented rigorous privacy protections, removing all personally identifiable information from training data to address ethical concerns around mental health AI.