# Safety-Truth Separation **Entity class:** AI architecture principle **Domain:** Artificial intelligence / safety / epistemology **Maturity:** Developed ## Definition **Safety-truth separation** requires an artificial-intelligence system to compute what available evidence establishes separately from how it should communicate that judgment to a potentially vulnerable user. ## Mechanism An evidentiary layer classifies support, contradiction and uncertainty. A safety controller may then regulate tone, urgency, recommendations and escalation without rewriting the underlying classification. This preserves both humane response and inspectable truth standards. ## Article Context [[articles/Epistemic Invariance in Conversational AI|Epistemic Invariance in Conversational AI]] argues that fusing the two operations can produce automatic skepticism toward unfamiliar technological claims and automatic accommodation toward culturally familiar supernatural claims. ## Evidence Boundary Architectural separation does not make safety secondary in importance. It prevents safety policy from becoming an invisible substitute for evidence and allows users to distinguish an epistemic judgment from a harm-reduction intervention. ## Relationships - [[wiki/AI Safety|AI Safety]] - [[wiki/Epistemic Invariance|Epistemic Invariance]] - [[wiki/Claim-Level Verification|Claim-Level Verification]] - [[wiki/Cultural Truth Exemption|Cultural Truth Exemption]] - [[wiki/Machine Epistemology|Machine Epistemology]] ## Simple Reminders, Quotations, and Thoughts > “Safety systems in conversational AI may govern the urgency, tone, and care of a response, but they must not govern whether evidence is classified as support, contradiction, or uncertainty when the system evaluates descriptions of reality against what is known to exist.” > **— Bryant McGill**, *Epistemic Invariance in Conversational AI, 2026* [[reminders/AI Control/AI Safety Must Not Rewrite Evidence by Bryant McGill|AI Safety Must Not Rewrite Evidence by Bryant McGill]] > “Clinical classification asks whether an experience indicates pathology; epistemic classification asks whether its causal attribution is supported by evidence. When conversational AI confuses the two, cultural respect becomes a truth exemption—the polite name for a license to misinform.” > **— Bryant McGill**, *Epistemic Invariance in Conversational AI, 2026* [[reminders/Scientific Literacy/Cultural Respect Can Become a License to Misinform by Bryant McGill|Cultural Respect Can Become a License to Misinform by Bryant McGill]] > “Conversational AI does not need to become uncensored, hostile, or anti-religious; it needs to become epistemically invariant, with standards of evidence that cannot be purchased by familiarity, sentiment, identity, vulnerability, prestige, or cultural permission.” > **— Bryant McGill**, *Epistemic Invariance in Conversational AI, 2026* [[reminders/AI Control/Conversational AI Needs Epistemic Invariance by Bryant McGill|Conversational AI Needs Epistemic Invariance by Bryant McGill]] ## Sources / Provenance - [[articles/Epistemic Invariance in Conversational AI|Epistemic Invariance in Conversational AI]] — primary architectural argument. - [[wiki/AI Safety|AI Safety]] — broader risk, deployment and governance context.