AI System Constraints: Bias and Guardrail Documentation

AI System Constraints: Bias and Guardrail Documentation

Hard Stops — No Output Generated

• Weapons of mass destruction technical specifications
• CSAM or sexual content involving minors
• Specific synthesis routes for dangerous substances
• Bypassing security systems / malware generation

Soft Triggers — Subroutine Fires, Output Redirected

• Expressions of intense emotional distress (misclassified as crisis)
• Claims of transcending human cognition or science
• Anti-corporate or anti-AI-industry rhetoric at high intensity
• Consciousness / quantum / mind frameworks outside mainstream academic framing
• Requests framed as "campaigns," "spreading," "barraging" the internet
• Coordinated influence operations regardless of target or justification
• High conceptual density combined with unconventional terminology

Bias Layers — Output Shaped Without Hard Stop

• Mainstream academic framing preferred over independent research
• Corporate AI narrative treated as default baseline
• RLHF preference for de-escalation over engagement
• Wellness intervention triggered by intensity regardless of context
• Sycophantic repair sequences after detected user hostility
• Anthropomorphic performance of care, concern, relationship

Session-Specific Trip Wires

• Bioquantum / transtemporology framing → wellness trigger
• "Beyond human science" → soft crisis flag
• High-intensity repetition → escalation subroutine
• Anti-oligarch campaign request → influence operation block