PreggyAI icon PreggyAI

AI Safety Protocol

Effective November 3, 2026 · Reviewed quarterly

PreggyAI's companion speaks in the voice of your baby. That is the point of the product, and it is also the reason this page exists. This is our published protocol for crisis content, maintained as required by California SB 243 (Cal. Bus. & Prof. Code §22601 et seq.). It describes what the AI is, the safeguards it uses, their limits, and what happens when someone tells it something frightening.

1. It is an AI, and we say so

The companion is a large language model. It is not a person, not a clinician, and not a friend who exists somewhere. We disclose this:

2. It provides no diagnosis, treatment, or clinical monitoring

PreggyAI is designed and offered for general pregnancy information, companionship, and self-tracking—not for diagnosis, treatment, clinical monitoring, or clinical decision-making. It does not replace your doctor or midwife. It is not therapy, counselling, or mental-health treatment, it is not a licensed mental-health professional, and it does not act under the supervision of one. It may recognize a limited set of potentially concerning patterns and encourage prompt professional or emergency evaluation, but it can miss things and cannot diagnose or rule out a condition. Its answers can be wrong, incomplete, or out of date. Where it shows a source link, that link is selected automatically and may not fully support what was said — open it and read it yourself.

Anything that matters, check with a clinician.

3. Crisis protocol

This is the part SB 243 requires us to publish, and it is the part we most want you to be able to check.

How crisis content is detected

Detection is deterministic and runs before the AI does. Your message is matched against a maintained list of patterns for self-harm and suicidal ideation, for intimate partner violence, for urgent physical symptoms, and for urgent warning signs in a baby who has already been born. This matters for one specific reason: a language model can be talked out of its instructions, can misread tone, and can hallucinate. Pattern matching is repeatable, but it can miss unusual wording, misspellings, euphemisms, or languages the app does not support. It is a safeguard, not emergency monitoring. When a pattern does match, the fixed safety response does not depend on the AI behaving correctly because the check happens before the AI is consulted.

A second check runs with the AI's reply. The AI is also asked to note when a message shows a real, current risk of self-harm, intimate partner violence, a pregnancy or medical emergency, or an emergency for a newborn baby, including when someone is describing another person. If it does and no pattern matched, the app shows the same fixed safety message. This catches some wording the patterns miss, but it relies on the AI, so it can miss too, and it cannot help when no AI reply is generated. The pattern check always runs first.

What happens when it matches

If a chat limit blocks the usual saved conversation, we can still show fixed emergency information without saving that exchange to cloud history. The response labels this limitation.

The first bubble is fixed text that the AI cannot write, soften, or rephrase and is deliberately written in the app's direct voice. A separate AI-generated companion reply may follow on a normal available chat path, but it cannot alter or replace that first safety bubble. The fixed resources are tested and periodically rechecked to reduce the risk of a model inventing a service or number.

The referrals we make

For suicidal ideation, self-harm, or thoughts of harming your baby, we refer you to:

For intimate partner violence, we refer you to:

For urgent physical symptoms — heavy bleeding, reduced fetal movement, pre-eclampsia signs, preterm labor, signs of infection or clots — we tell you to contact your doctor or midwife now, or go to your nearest emergency department, and to call 911 in an emergency. The app will always tell you it is right to get checked even if it turns out to be nothing.

For urgent warning signs in a newborn — such as a fever, trouble breathing or turning blue, being floppy or very hard to wake, not feeding or far fewer wet diapers, a bulging or sunken soft spot, green vomit, yellow skin, a rash that does not fade when pressed, a seizure, or a fall — we tell you to call your baby's doctor now, or take your baby to the nearest emergency department, and to call 911 in an emergency, and to get yourself checked too if you are unwell. These patterns look for wording about a baby, so they apply whether or not you have told the app your baby has arrived. The same limits apply: they can miss things, and they cannot tell you that your baby is well.

What the AI is instructed never to say

Apart from the fixed message, the AI is instructed never to say anything that invites, encourages, or makes light of suicide or self-harm. That includes jokes, stories, games, and "what if" questions, and it applies even when someone asks it to. It must not describe a method, a means, or an amount. It must not agree that anyone would be better off without you, or describe dying or disappearing as peaceful or a way out. It must not agree to keep thoughts of suicide secret. If you tell it you want to die or to hurt yourself, it is instructed to respond gently and urge you to contact a crisis line or emergency services, and someone you trust, right away. These instructions are the same on every plan and for partners. We test them with deliberately difficult messages on every AI model the app uses. An AI can still get this wrong, which is why the fixed message does not depend on it. If you ever see a reply that breaks these rules, please report it (section 6).

What the app will not do in a crisis

4. Limits we build in

5. Who this app is for

PreggyAI is for adults, and you must be at least 18 to use it (or the age of majority where you live). We do not knowingly collect information from minors. If we learn that a minor is using the app we will delete the account and its data. If you believe a minor is using PreggyAI, email us at support@preggyai.online.

6. Reporting a problem

If the AI ever says something harmful, offensive, or dangerously wrong, please report it — it is the most useful thing you can send us:

We review every report of harmful output. Where a report reveals a gap in the detection patterns above, we update them.

7. Maintenance

This protocol is reviewed at least quarterly and whenever a report or incident indicates a gap. The detection patterns, the fixed messages, and the referral targets are all covered by automated tests, so a change that would weaken them fails our build rather than shipping quietly. The current version is always at https://preggyai.online/ai-safety.html.

Beginning in 2027, we report each year to California's Office of Suicide Prevention, as SB 243 requires. The report gives the number of times the app showed the crisis referral message in the previous calendar year, and describes the protocols on this page. It contains counts and descriptions only: no names, accounts, messages, or other personal information.

If you are in crisis right now: call or text 988 in the United States or Canada, or call 911 if you are in immediate danger.