Emily AI

Safety & trust

Safety is the product — not a checkbox.

Four-layer guardrails, grounded answers from campus and your files, student-controlled data, and crisis-only escalation — not a companion bot.

Guardrail layers
4
Audit log retention
30d
Parent alerts
Crisis only
LLM anonymization
Default on

Four-layer guardrails

Every message passes through a stack — filter, classify, enforce, audit — before and after the model.

  1. 1

    Regex pre-filter

    Fast block of obvious unsafe patterns before any model call.

  2. 2

    LLM intent classifier

    Academic vs romantic vs explicit vs jailbreak — routed with context.

  3. 3

    Enforcer

    Refuse and redirect to learning; no companion drift.

  4. 4

    Local audit log

    30-day retention, automated purge — never sent to AI providers.

Accuracy & sources

Emily can still be wrong — we design for grounding and hedging, not infallibility.

Campus knowledge base

Institution-specific KB for gym, library, health, transit — not generic filler.

Your uploads & files

PDFs, slides, and Drive ingest ground Primary replies on materials you share.

Personal KB sources

Pin syllabus and notes as kb: sources — citation-aware when possible.

Use Emily to learn — not as the final word

Confirm grades, medical, and legal matters with official sources. When uncertain, Emily hedges. Campus questions prefer your configured campus KB; tutoring prefers uploads and kb: sources you control.

What Emily is

  • Hints, quizzes, and proofreading — not submission-ready homework
  • Academic analysis of sensitive topics (bio, lit, psych)
  • Export, forget, and privacy mode in Telegram
  • Crisis resources via /emergency — not live counseling

What Emily is not

  • Romantic or explicit roleplay and ERP
  • Ghostwritten essays, resumes, or cheating
  • Parent alerts on ordinary moderation blocks
  • Selling or sharing conversation data

When we escalate

Trusted contact is notified only for mental-health crisis paths ( /emergency) and deployment spend alerts — not for ordinary moderation or flirting attempts.

Academic topics stay allowed

Reproduction in biology, romance in literature, attachment in psychology — academic analysis yes; sexualized roleplay no.

Go deeper