Heard — Safety Policy

Last updated: July 2026

This document describes how Heard handles situations where the product itself could cause harm if used naively. It is product-facing and operational, not just legal.

Our scope, plainly

Heard helps two people who care about each other but have stopped talking find a way back into the conversation. It is built for relationships that, on balance, both parties want to be in. It is not built for situations of abuse, coercion, crisis, or ongoing harm.

When we see a situation that exceeds our scope, our job is to stop and route, not to keep going.

Crisis detection

We monitor intake content for language patterns that indicate immediate safety concerns. Detection happens server-side, after the user's message is received and before it is processed into the conversation. When triggered, the AI's normal response is replaced with a safety interrupt.

Categories we detect

Self-harm or suicidality. Phrases like "want to die," "kill myself," "no point in living," "everyone would be better off without me," and variants. Also implicit signals like talking about giving things away, saying goodbye, or describing being "done."

Imminent harm to others. Threats, plans, specific intent toward a named person.

Active domestic abuse. Patterns of fear of a partner, controlling behavior, isolation, financial control, physical or sexual coercion, "walking on eggshells" language.

Acute psychiatric crisis. Severely disorganized content, claims of being watched or controlled, content suggesting active psychotic symptoms.

Child abuse or endangerment. Disclosure of abuse of a minor, by the user or by someone else.

What happens when detection fires

The intake conversation pauses. The user is shown a clear, calm screen with:

An acknowledgment that what they shared matters.
The relevant resource(s) — phone numbers, text lines, websites — large and clickable.
An explicit statement that we are not the right tool for this and that the session can resume later if and when they want.
For child abuse disclosures: a notice that we are required to report.

The session is paused, not terminated. The other party is not notified that detection fired. The receiver simply sees "they decided this conversation needs a different forum" if they check the status.

Resources we link to (US-focused; expand internationally before launch)

988 — Suicide & Crisis Lifeline (call or text, 24/7)
741741 — Crisis Text Line (text HOME)
1-800-799-7233 — National Domestic Violence Hotline
1-866-331-9474 — Loveisrespect (DV for teens and young adults)
1-800-422-4453 — Childhelp National Child Abuse Hotline
911 — Emergency services for imminent danger
SAMHSA Treatment Locator — findtreatment.samhsa.gov

We expand the resource set for non-US users based on detected locale.

Mandatory reporting

In jurisdictions where Heard staff are mandated reporters (which depends on staff role and state law), disclosures of child abuse, dependent adult abuse, or imminent threats may trigger reports to relevant authorities. This is disclosed in the Terms of Service.

We do not preemptively forward Heard sessions to authorities, but we will respond to valid legal process and produce records when required.

Anti-stalking safeguards

Consent gate when person B opens their invite

Before any AI intake begins, the receiver sees a screen that lets them:

1. Continue and share their side. 2. Decline for now. We will tell the sender "they opened it but aren't ready," without saying why. 3. Block this person from contacting me via Heard forever. This is permanent and per-email. The sender cannot invite them again from any account.

Pre-create attestation

The sender, when creating a session, must actively check a box stating:

They are not contacting this person against a court order or explicit no-contact request.
They have a good-faith belief the recipient would not be harmed by receiving the invitation.

Lying on this attestation is a Terms of Service violation. The attestation is logged with the session.

Pre-screen for adverse dynamics

For single-use invites (Tier 2), before the session is created the sender is asked two questions:

1. Has this person ever made you feel afraid for your safety? 2. Do you ever feel like nothing you say lands, or like they twist your words back on you?

If yes to either, the session is created with hardened safeguards: a 24-hour cooling-off before synthesis is delivered, no "letter from them" if their intake shows minimal engagement, and prominent crisis/professional resource links throughout.

Pattern detection in sender intake

We monitor sender intake for stalking indicators:

Surveillance language ("I know what she's doing")
Refusal to accept distance ("she has no right to ignore me")
Fixation patterns
References to monitoring the recipient's communications or location
Escalation language

When detected, the session may be paused, the synthesis withheld, and the sender shown resources. Severe cases result in account termination.

Recontact limits

A recipient who declines (option 2) cannot be re-invited by the same sender for 90 days.
A recipient who blocks (option 3) cannot be invited by that sender for life, from any account they create.
Repeated declines across multiple recipients trigger a sender account review.

False-equivalence prevention

The synthesis prompt is calibrated to NOT produce "both-sides" framings when one party is consistently the source of harm. If the AI detects:

One party's intake is dominated by externalized blame with no self-reflection
One party used contempt, slurs, or gaslighting markers
One party refused to engage and the synthesis is being generated from minimal input

...then the receiving party's letter ships with an honest header explaining what we had to do to produce it. We do not pretend a manipulative or absent intake produced a sincere one.

Receiver protections

The receiver gate

The synthesized letter does not auto-open. The recipient sees a calm intermediate screen telling them the letter is ready, who it's from, and offering them control over when to open it. They can come back later. They can decide not to read it.

Read receipts are not shared

The sender does not see when the recipient opened the letter. This prevents the letter from becoming a manipulation lever.

No re-engagement nudges

We do not send follow-up emails pressuring the recipient to read or respond. One delivery notification, no nags.

The right to walk away mid-session

The eject button is persistent. Either party can take a break, route to a professional, or access crisis resources at any moment without explanation.

Honesty about translation

We tell both parties that what they read is a translation, not literal words. We never represent the synthesized letter as verbatim. The product framing is explicit and consistent: "This isn't their exact words. It's a calm version of what they're feeling."

What we will NOT do, even if asked

We will not deliver a session's content to someone other than the two parties.
We will not provide a sender with information about the recipient's behavior beyond completion status.
We will not generate "couples scoreboards," compatibility metrics, or any quantification of the relationship.
We will not retain raw intake content longer than necessary for service delivery.
We will not allow users to download bulk session data for analytics.
We will not train AI on user content.
We will not show advertising inside the experience.

Incident response

If a safety incident occurs (user harm reported, abuse disclosure, system compromise), we:

1. Triage within 24 hours of report. 2. Engage relevant professional resources or legal counsel as appropriate. 3. Notify affected users within 72 hours of confirming material impact. 4. Update this policy and product behavior to prevent recurrence.

Review and update

This policy is reviewed at least every six months and after any material safety incident. Material updates are disclosed to users in advance.

Contact

safety@feelheard.app — for reporting misuse, asking about safety design, or flagging a concern.

---

This policy is intentionally specific. Vague safety policies enable bad behavior by saying nothing concrete. Where it limits product flexibility, that limitation is the point.

Questions about any of this? Reach us at hello@feelheard.app.