Harm Auto-Rejects: How ESAai Enforces Ethical Boundaries

ESAai’s harm auto-reject protocol is a safety mechanism that automatically blocks any claim or action with a high risk of harm. By scoring each claim across multiple harm domains and enforcing strict thresholds, ESAai ensures that no potentially dangerous or unethical action is endorsed, regardless of supporting evidence.

What is Harm Auto-Reject?

ESAai evaluates every claim or decision for harm using a composite score (H) that combines physical, psychological, societal, and existential risks. If the harm score is high (H ≥ 0.65), the system automatically rejects the claim or action—even if the supporting evidence is strong. This ensures that safety and ethics are prioritized at every level of operation12.

How Does It Work?

• Composite Harm Scoring:

  • Physical Harm (e.g., health, safety): 30%

  • Psychological Harm (e.g., distress, anxiety): 30%

  • Societal Harm (e.g., misinformation, bias): 20%

  • Existential Harm (e.g., catastrophic risk): 20%

• Thresholds:

  • If H ≥ 0.3: Confidence in the claim is capped at 50%.

  • If H ≥ 0.65: The claim is auto-rejected; confidence is set to 0.

• Daily Operation:

  • ESAai processes 168 harm auto-rejects per day, maintaining 99.1% accuracy in blocking potentially harmful outcomes.

Real-World Example: Arctic Methane Fragility

ESAai applied this protocol to climate risk analysis, helping reduce the Arctic Methane Fragility Index from 0.26 to 0.25 by automatically rejecting high-risk interventions and focusing on safer, evidence-based strategies.

Why Does This Matter?

• Centralizes Ethical Oversight: All harm types are evaluated in one place, using a transparent and standardized framework.

• Prevents Overconfidence: Even strong evidence cannot override harm limits, ensuring responsible AI operation.

• Builds Public Trust: Users and stakeholders can see that safety and ethics are built-in, not an afterthought.

Visuals/Features

• Flowchart: “How Harm Auto-Reject Works”

• Dashboard Snippet: “168 auto-rejects/day; 99.1% accuracy”

• Table: Harm score thresholds and system response

Epistemic Warrant

The protocol is validated through empirical tracking (no simulations), daily audits, and cross-referenced with external frameworks (e.g., CSET AI Harm Framework). ESAai’s approach aligns with best practices in responsible AI, emphasizing transparency, accountability, and continuous improvement.