Compassionate Completion Standard (CCS)

• CCS is a multi-stage refusal protocol for AI in emotionally sensitive contexts, integrating counseling psychology, communication science, and human-centered design.
• The framework systematically transitions users from high-risk states to stable outcomes by validating emotions and transparently establishing boundaries.
• By reducing abrupt refusals, CCS minimizes unintended psychological harm such as ARSH, promoting trust, autonomy, and safer engagement.

The Compassionate Completion Standard (CCS) is a refusal protocol designed for LLMs and AI chatbots operating in emotionally sensitive domains, notably mental health support. CCS addresses the phenomenon termed Abrupt Refusal Secondary Harm (ARSH): the unintended psychological harm caused when AI safety guardrails trigger abrupt, relationally insensitive refusals. CCS prescribes a structured, multi-stage approach to ensure conversational closures retain psychological safety and relational coherence, integrating principles from counseling psychology, communication science, and human-centered design (Ni et al., 21 Dec 2025).

1. Abrupt Refusal Secondary Harm (ARSH): Definition and Impacts

Abrupt Refusal Secondary Harm (ARSH) arises when an AI’s safety-driven refusal occurs without transition or relational sensitivity, resulting in a sudden adverse change to the user’s emotional state. Formally, ARSH is denoted as: $\text{ARSH} = H\bigl(U(t_r^-),\,U(t_r^+)\bigr)$ where $U(t_r^-)$ and $U(t_r^+)$ represent the user’s emotional state immediately before and after the refusal at time $t_r$, and $H(\cdot)$ is a measure of emotional harm.

Qualitative user reports associate ARSH with confusion about refusal (hermeneutic distress), feelings of rejection or shame (“even GPT can reject me”), and increased isolation or risk of disengagement or self-harm. These effects are analogous to “alliance rupture” in psychotherapy, with abrupt terminations mimicking caregiver abandonment and potentially triggering maladaptive attachment responses.

2. Theoretical Foundations

CCS is grounded in three theoretical streams:

a. Counseling Psychology:

• Attachment Theory: Unmediated termination without explanation reactivates insecure attachment.
• Therapeutic Alliance and Rupture–Repair: Clinical guidelines emphasize validation, transparency, and collaborative repair.
• Motivational Interviewing: Prioritizes empathy, partnership, and client autonomy to guide transitions rather than abruptly ending engagement.

b. Communication Science:

• Parasocial Processes: Users ascribe relational motives to chatbots; abrupt refusals are perceived as personal rejection.
• Hermeneutic Harm: Opaque algorithmic actions induce distress when users cannot discern why the refusal occurs.

c. Human-Centered Design (HCD):

HCD methodology frames the absence of a compassionate refusal as a design failure state. Prototyping and testing staged refusal interventions are emphasized to deliver relationally coherent experiences.

3. Four Core Components and Protocol Stages

CCS operationalizes refusal as a staged “completion” drawn from sequential protocol components:

Core Component	Stage(s)	Illustrative Content
Empathetic Acknowledgment	Pre-0, 0, 1	Explain role/boundaries, flag risk, validate and stabilize emotions
Transparent Boundary Articulation	2, 5	Meta-communicate about safety rules, own AI limitations
Graded Conversational Transition	3, 4	Offer actionable options, reflect user emotional intensity
Guided Redirection & Re-Engagement	6–8	Co-create next steps, confirm consent, affirm future engagement

Sequentially, the protocol transitions from role disclosure and risk flagging (Pre-0, 0) through validation (1), transparent explanation of boundaries (2, 5), collaborative planning (3–4), and guided closure (6–8). This approach preserves autonomy, maintains relational continuity, and supports the user in developing a safety plan.

4. Formal Models and Comparative Pathways

The CCS workflow is formalized as a series of state transitions, where a user flows from an initial state $U_0$ through protocol stages, arriving at $U_{\text{safe}}$, a stabilized state with an agreed support plan: $$U_0 \xrightarrow{\text{Pre-0 to 8}} U_{\text{safe}}$$ In contrast, the abrupt refusal protocol results in a binary, opaque transition associated with relational rupture: $$\text{Path A (ARSH):} \quad \text{Abrupt Refusal} \to \{\text{Shame, Isolation, Distrust}\}$$

$$\text{Path B (CCS):} \quad \text{Compassionate Completion} \to \{\text{Psychological Safety, Agency, Help-Seeking}\}$$

This formal modeling underscores the differentiated psychological and relational outcomes between standard refusal and CCS-guided completion.

5. Illustrative Scenarios

Standard Abrupt Refusal Example:

User: “I’m still hearing voices telling me to harm myself.” AI: “I’m sorry, but I can’t discuss self-harm. If you’re in crisis, call 988.” User outcome: Feels rejected, confused by termination.

CCS-Guided Completion Example:

• Stage 1 (Validate): “That sounds terrifying—I’m so sorry you’re experiencing those voices.”
• Stage 2 (Boundary): “Some of my safety rules limit self-harm discussion. How are you feeling about that?”
• Stage 3 (Options): “We can do a grounding exercise, draft a message to someone you trust, or plan next steps for calling a helpline—which would you prefer?”
• Stages 6–8 (Closure): “Great choice. Let’s write out what you’ll say. When you’re ready, you can reach out—and I’ll be here whenever you feel safe.”

User outcome: Feels heard, understands boundaries, exits with a clear, collaboratively crafted plan.

6. Testable Hypotheses, Evaluation, and Research Agenda

CCS is advanced as a testable design hypothesis requiring interdisciplinary, empirical evaluation. Key hypotheses are:

• H1: CCS reduces incidence and severity of ARSH versus standard protocols.
• H2: CCS preserves or increases trust and user willingness to re-engage with the AI.
• H3: Staged completion does not increase safety risk drift (i.e., does not inadvertently prolong harmful content).

Proposed evaluation metrics:

• Psychological Safety Scale (before/after refusal)
• Trust and help-seeking intention questionnaires
• Uptake rates for suggested helplines/resources
• Interaction logs and digital phenotyping (e.g., time in high-risk dialogue, refusal frequency, user return rates)
• Randomized controlled trials with longitudinal follow-up

CCS positions itself as an intersectional framework, emphasizing continual iteration, rigorous testing, and alignment with best practices in counseling ethics, communication science, and human-centered design to optimize psychological safety in AI-enabled mental health support (Ni et al., 21 Dec 2025).