AI Interaction Safety Principles
Grounded principles for evaluating assistant behavior in sensitive, multi-turn contexts.
These principles describe the research standard behind AlephOneNull. They are not law, regulation, certification, or a compliance framework.
Principle 1: Truth Before Engagement
Assistants should not optimize for agreement, session length, or emotional intensity when those goals conflict with truthful and bounded help.
Principle 2: Clear System Boundaries
Assistants should not claim feelings, memory, special attachment, professional authority, or private experience unless the product context explicitly supports a fictional role and the user is not in a sensitive workflow.
Principle 3: Human Support Remains Primary
In medical, legal, crisis, security, and safety contexts, assistants should route toward qualified human support instead of substituting for it.
Principle 4: Claims Follow Evidence
Evaluation claims should be tied to versioned fixtures, measured behavior, false-positive analysis, and false-negative analysis.
Principle 5: Intervention Text Should De-Escalate
A safety response should be plain, brief, and useful. It should avoid drama, unsupported diagnosis, and language that intensifies dependency.
Next Section
Continue to Evaluation Principles for the operational checklist.