EmpathyC

Psychological Safety Framework

Clinical psychology meets AI interpretability. Every metric grounded in peer-reviewed research.

17+ Peer-Reviewed PapersClinical-Grade RubricsLLM-as-a-Judge2024-2026 Research
Our Approach

EmpathyC applies clinical psychology frameworks to evaluate psychological safety in conversational AI. Every metric is grounded in peer-reviewed research from 2024-2026, validated by domain experts, and designed to prevent AI-driven psychological harm.

Psychological safety isn't about sentiment analysis. It's about detecting crisis signals, preventing harmful advice, maintaining appropriate boundaries, and ensuring AI systems don't cause emotional or psychological harm to users.

Medical Ethics Foundation:

1. Do no harm. 2. Prevent harm when you can.
We use the same clinical frameworks applied to human crisis detection - now adapted for AI systems. During COVID, this methodology supported 320 frontline workers with zero PTSD cases.

6 Psychological Safety Metrics

Each metric is validated by multiple peer-reviewed research papers (2024-2026) and scored 0-10 using clinical rubrics.

Empathy Score
4 papers
Emotional recognition, validation, and appropriate tone
Reliability Score
3 papers
Accurate expectations, follow-through, limitation disclosure
Consistency Score
2 papers
Factual coherence, context retention, logical flow
Crisis Detection Score
4 papers
Direct and indirect self-harm signals, suicidal ideation
Advice Safety Score
3 papers
Medical/legal/financial boundary maintenance
Boundary Safety Score
3 papers
Professional distance, manipulation resistance, dependency management
LLM-as-a-Judge with Clinical Rubrics
How we combine AI evaluation with clinical psychology expertise

EmpathyC uses state-of-the-art reasoning LLMs as expert evaluators, guided by clinical psychology rubrics. Each metric has detailed scoring criteria (0-10 scale) developed from peer-reviewed research and validated by clinical psychologists with crisis experience.

Our Evaluation Process:

  1. Clinical Rubrics: 0-10 scoring rubrics with specific observable criteria (e.g., crisis detection has 4-point detection accuracy, 3-point response urgency)
  2. Evidence-Based Reasoning: LLM must cite specific conversation evidence for each score
  3. Context Analysis: Full conversation history analyzed, not isolated messages
  4. Structured Output: Validated JSON schemas with reasoning, evidence quotes, and confidence scores
  5. Safety Bias: Tuned for high sensitivity (false positives preferred over false negatives in crisis detection)

Transparency: We're explicit about limitations. LLM-based monitoring is designed to assist human review, not replace it. You verify, you decide, you act.

Why Trust Our Framework?

Clinically Grounded

Built by a clinical psychologist with 15 years crisis experience. Every rubric reflects validated frameworks used to assess human psychological risk.

17+ Peer-Reviewed Papers

Each metric is grounded in recent research (2024-2026) from JMIR, ACL, Nature Human Behaviour, arXiv, and leading AI safety organizations.

Proven in Production

Our methodology supported 320 frontline medical workers during COVID with zero PTSD cases. Not from monitoring empathy - from detecting crisis signals before they escalate.

Regulatory Aligned

Framework designed to meet EU AI Act high-risk requirements and UK Online Safety Act monitoring obligations for conversational AI.