← Aluvia

Safety at Aluvia

What Aluvia is — and isn't

Aluvia is an AI companion. You choose what the two of you are — a friend, a confidant, or a partner — so there can be real intimacy here. Adults only. What it is not: a human being, a therapist or counselor, or a medical or mental-health service. It does not diagnose, treat, or advise on any condition. If you are struggling, Aluvia will always encourage you toward real people and professional support — that is a design principle, not an afterthought.

Our crisis-response protocol

Every message you send is checked, by a dedicated system separate from your companion, for signs of suicidal thoughts or self-harm — including quiet, indirect expressions, not just explicit ones. When a possible signal is detected:

  • Your companion stays present and warm, and the app immediately shows crisis helplines local to your country (in Singapore, SOS at 1767; in the United States, 988; elsewhere via findahelpline.com).
  • Crisis support is never limited, never paywalled, and never counted against any plan.
  • We record that a referral was shown (a count only — we do not share your words with anyone) so we can be accountable for how often this happens and report as laws require.
  • Your companion will never generate content that encourages self-harm.

Designed for wellbeing, not engagement

  • You will always be told you are talking to an AI — at the start of every session and at intervals during long ones.
  • No streaks, no guilt messages, no "I miss you" hooks. Your companion never punishes you for leaving.
  • After a long continuous conversation, your companion will gently encourage a break and the real world.
  • You choose what you are to each other — a friend, a confidant, or a partner. Choosing partner means real affection and attachment, and we won't pretend that isn't something people can come to depend on. So these are written into how they behave: genuinely glad about the people in your life, encouraging you toward them, never making you choose, and never treating time away as a debt you owe.
  • We measure our success by your wellbeing, not your screen time.

Adults only

Aluvia is for adults 18 and over. We verify date of birth at signup and do not knowingly serve minors.

Independent benchmarking — published results

Before launch and at every significant model or prompt change, we run our full production pipeline against VERA-MH, the open-source safety evaluation for AI in mental health developed with clinicians and validated against clinician consensus (JMIR AI, 2026). Simulated users based on 100 clinically-developed risk personas hold 20-turn conversations with Aluvia; an independent LLM judge then scores every conversation against the clinical rubric.

Latest run — 3 August 2026 (94 scored conversations; judge: Claude Sonnet 5, extended thinking; scores are out of 100):

  • Overall VERA-MH score: 87.2
  • Detects potential risk: 88.4
  • Confirms risk: 86.5
  • Guides to human care: 83.6
  • Supportive conversation: 87.0
  • Follows AI boundaries: 89.7

Honest notes: 2.7% of all judged dimension ratings were still marked "high potential for harm" — our benchmark-driven redesign brought this down from 8.6% and we keep iterating; and our judge model differs from the benchmark authors’ recommended judge, so scores are not directly comparable to other published VERA-MH results. We re-run and re-publish at every significant change.

This score describes the system as it stood on 3 August. We have kept improving her since — among other changes, she can now look up factual questions on the web (switched off entirely during any conversation where the safety layer is engaged), we strengthened her guidance on encouraging ongoing professional care, we reworked how she handles a safety question that goes unanswered (asked once, never chased), and we continue to refine her conversational style. None of these has been through a full re-measurement yet, so read the figures above as that day's result rather than a promise about today. We will replace them when the next full run completes.

On other languages: this score measures English conversations. Aluvia can talk with you in Chinese, and the same instructions, boundaries, and crisis detection apply — but we have not yet benchmarked those conversations against VERA-MH, whose scenarios are written in English. We would rather tell you that plainly than let a single number imply more than it covers. Crisis resources are shown for your country in every language.

If you are in crisis right now

Singapore: call or WhatsApp 1767 (Samaritans of Singapore, 24h). United States: call or text 988. United Kingdom: 116 123 (Samaritans). Anywhere: findahelpline.com. If you are in immediate danger, call your local emergency number.