AI in pregnancy apps: what is actually shipping

It is 11 p.m. and you are asking ChatGPT about a symptom

You are 28 weeks pregnant. The toddler is asleep, the partner is asleep, the bedside light is the only one still on, and you have a question about a tightness across your jaw that has been there since dinner. You do not want to call the after-hours line. You do not want to wake the GP. The browser is already open on your phone, and so, increasingly, is a chat window.

Most pregnant women you know do this now. The chatbot is open. The question is honest. The answer comes back in a confident tone, regardless of whether it is right. There is a version of the story that gets shared every few months in the press, the one where the chatbot caught something a clinician would have caught at the next visit. That story exists. It is a thin slice of the picture.

The honest version is harder. ChatGPT does not know your pregnancy week, your appointments, your medication list, your mood trend, or whether you have just walked up two flights of stairs. It is a smart stranger you have never met, briefed only by your prompt at 11 p.m.

The pregnancy app on your phone, by contrast, has all of that context, in theory. This article is about what that “in theory” actually looks like in May 2026. What AI features have shipped across the major pregnancy apps. What they do well, what they do badly, and where myCocoon’s choices put it on a different line than most of the category.

What is actually shipping in pregnancy-app AI today

A useful thing to do, before having a strong opinion, is to look at what is on the App Store. Of nine major pregnancy and fertility apps surveyed in May 2026, five publicly advertise a generative-AI feature inside the app. Four do not advertise one at all.

• Flo advertises an in-app health assistant. A 2025 Databricks press release confirms Flo fine-tunes its own large language models on Mosaic AI, but the model vendor is not publicly named. Flo is also the company the FTC settled with in 2021 over sharing fertility-tracker data with Facebook and Google analytics, and a 2024 class action added another 56 million dollars in settlements. (FTC, 2021, Databricks, 2025)
• Glow Nurture rebranded its App Store listing to “AI Pregnancy App” and markets unlimited access to a generative assistant called GlowGPT plus a library of medically reviewed articles. Model vendor not disclosed.
• Ovia Health, distributed largely through employer benefits, runs a 24/7 chatbot and voice-input symptom tracking. Model vendor not disclosed.
• The Bump, owned by The Knot Worldwide, ships an Ask AI Q&A feature inside its pregnancy tracker. Mozilla’s “Privacy Not Included” project flagged The Bump for broad data collection, the use of data resellers, targeted ads, and possible sale of personal information. (Mozilla, “Privacy Not Included”)
• Premom ships an Ask AI feature whose privacy disclosure names OpenAI as a downstream vendor. Premom previously settled with the FTC over data-sharing without consent.

The other four, What to Expect, BabyCenter, Pregnancy+ from Philips Avent, and Hello Belly, do not advertise a generative-AI assistant at the parent-app level as of this writing. None of the five vendors that do advertise an assistant publicly discloses a few things you would want to know before you typed: which model is answering, how long your prompts are kept, and whether your prompts are used to train future models.

That last point is not a niche concern. The consumer versions of ChatGPT, Claude, and the public Gemini app are not covered by HIPAA Business Associate Agreements; only the enterprise tiers and the first-party APIs are. (HIPAA Vault) When a pregnancy app routes your question to a model, the right question is not “is the model smart enough.” It is “which model, under whose contract, with what retention, and disclosed to me how.”

Where the accuracy actually lives, and where it leaks

The other thing worth knowing before relying on any pregnancy chatbot is how good the answers really are. The peer-reviewed picture is more nuanced than either the boosters or the cynics suggest.

A 2024 study in npj Women’s Health tested ChatGPT against the UK Royal College of Obstetricians and Gynaecologists exam. It scored 72.7 percent on basic obstetric medical knowledge and 50.4 percent on decision-making and clinical reasoning. (Hall et al., npj Women’s Health, 2024) A 2024 paper in the International Journal of Gynecology and Obstetrics evaluated ChatGPT on common obstetric questions and noted that the model “occasionally missed key clinical nuances.” (Peled et al., 2024) A 2025 evaluation in Frontiers in Artificial Intelligence of a retrieval-augmented pregnancy chatbot called DIAN found accuracy ranging from 83 to 89 percent on nutrition advice and 96 to 98 percent on red-flag symptoms, with comprehensibility and readability varying sharply by topic. (Dong et al., Frontiers in AI, 2025) The press version of those numbers tends to focus on the high end. A 2025 viral story described a third-trimester reader who asked ChatGPT about a tight jaw; the model suggested checking blood pressure, which turned out to be 200 over 146, severe preeclampsia territory. (Upworthy, 2025) That story is real. It is also the survivor case. The variance in the studies above is the picture you do not see in the press.

Two things follow from those numbers. First, the average accuracy is higher than the lay press tends to suggest. A pregnancy chatbot, set up with the right scaffolding, gets red-flag answers right in the high nineties. Second, the variance is high in exactly the wrong places. The same model that flags preeclampsia symptoms reliably can be wrong about nutrition or non-evidence-based exercise advice, and you cannot easily tell which question is which while you are typing. The 2024 academic review that looked across a wide pregnancy-app field flagged at least one app for advice like “skip dairy and eat green foods for calcium,” which is the kind of confidently wrong sentence that does not look wrong in a chat bubble. (PMC review of pregnancy apps, 2020)

The honest read on AI in pregnancy is not that it is bad. It is that accuracy depends on the topic, the prompt, the model, and the scaffolding around it, and that the patient at 11 p.m. has no way to know which of those is in play. That is not a model problem. That is a product problem.

What myCocoon actually does, in plain language

myCocoon ships AI features. It would be dishonest to write a post about AI in pregnancy apps and pretend otherwise. The interesting question is which AI, where, and on what terms.

There are three layers, and they are arranged so the default is the quiet one. (For the full architectural picture, see the post on how myCocoon reads what is already on your phone and the comparison with Flo.)

The first layer runs on your phone, not in a cloud. Apple’s on-device language model, the one that lives in iOS 26, handles the small everyday work: cleaning up a voice-dictated journal entry, drafting a gentle prompt for the day, generating a short caption for a bump photo, narrating the chapters of your pregnancy story for the keepsake view. None of that text leaves the phone. The story narration, in particular, is on-device by design and not negotiable. Pregnancy data does not get to be the price of a feature.

The second layer is a rule engine. The cross-signal reads described in the previous post on this blog, where the app watches your sleep average, your resting heart rate, your steps, and your mood together and surfaces a card when the combination changes, are not AI. They are simple thresholds and trend checks against your own rolling baseline. They do not need a model. They need to be careful.

The third layer is opt-in cloud AI, off by default. When you turn it on, and only then, your questions go to Google’s Gemini, with two separate switches: one for chat-style questions, one for quiet features like the mascot’s greetings. Before any question leaves the phone it passes through a cleaning step that simplifies every field. Names and journal text never go. Raw values from the Health app, the precise heart rate readings, the exact weight, the timestamps, are stripped. What replaces them is rougher: your mood becomes one descriptive word (improving, stable, declining); your symptoms become a count rather than a list of names; your sleep, resting heart rate, steps, and activity become a daily band (last night’s sleep was short, typical, or long; your resting rate is in your usual range or above it; your day has been still, light, or active). Your pregnancy week, your trimester, and the question itself also travel. The privacy dashboard inside the app shows the exact thing that was sent, in the same words. No raw number ever leaves the device.

A separate safety check runs before any model call. If you type a phrase that matches a crisis pattern, the app responds locally with the appropriate resource and never sends the message to a model.

This is not the slogan version of privacy. When cloud AI is off, none of your data leaves the phone. When you turn it on, the boundary moves, and the new boundary is documented. Saying it any other way would be marketing, and the readers we built this for can tell.

What you have probably already noticed

If you have used a pregnancy app’s AI assistant in this pregnancy or your last one, you have probably noticed a few things.

The chat opens with an empty input box. It does not know which week you are in, or that your last appointment was a fortnight ago, or that you have logged restless legs three times this week. You type the question, the model answers, and the answer is generic, because the model has nothing else to go on. Or it answers in a tone that is uncannily warm in a way that does not feel like a person, because the system prompt is doing the heavy lifting and the system prompt is invisible to you.

You have probably also noticed the absence of a disclosure. Most pregnancy apps do not tell you, anywhere a normal person would look, which model you are talking to, what is sent, what is kept, or what is shared. The rare app that does tell you tends to bury it behind several taps. There is no industry norm for this yet. There probably should be.

You may have also noticed the second pattern, which is the chat’s reluctance to actually be a chat. Many of these assistants steer you, often correctly, towards “talk to your provider.” That is the safe default. It is also the default that makes the assistant near useless on the questions an assistant could actually help with: was today’s walk too far, is this pattern of restless legs the same one I had last week, am I drinking enough water given my activity level today. The narrowness is not because the model cannot answer. It is because the app has no other way to limit liability.

What you are noticing, in other words, is the gap between what is shipping and what would be useful. That gap is the design problem.

What myCocoon AI is, and what it is not

A few honest scoping notes.

It is not a medical device. myCocoon has no FDA clearance and no regulated medical-device designation. The app is consumer software with careful health-data handling. Anything that worries you on the screen is a reason to call your GP or midwife. The app’s job is to make that call easier, not to replace it.

It is not HIPAA compliant in the strong sense. The cloud AI path runs through Google’s AI service, which is governed by Google Cloud’s terms and not by a Business Associate Agreement. The defense, when cloud AI is on, is the cleaning step plus your consent, not a BAA. We tell you what is sent, you choose, and the raw numbers are stripped before send.

It does not know everything about you, on purpose. The cloud payload does not include your name, your partner’s name, your baby’s name, your appointment notes, the contents of your journal, or your raw vitals. That is a constraint, not an oversight. A model that is told less can be wrong about less.

Some of the AI features are still being built. Voice-mode, in particular, is the most explicit cloud surface the app has and the one we treat with the most care. The system instruction tells the model not to name itself or its vendor in conversation, the consent toggle is the same one that gates chat, and the payload is the same shape as the text path so a single privacy disclosure covers both.

It will not flatter you. The app is not designed to feel like a friend. It is designed to feel like a quiet adult in the room who has read your records, knows the week you are in, and notices when something has shifted. That is a different brief from the one most assistants are running.

Why the default is local, and why we still wrote this down

The reason cloud AI is off by default is not philosophical. It is practical. Most of the useful work in a pregnancy app is local work. The cross-signal reads do not need a language model. The journal cleanup runs faster on the phone than over a network. The story narration, written from your own week-by-week notes, is something we do not want to need a server for.

The reason cloud AI exists at all, behind a switch, is that some questions deserve a fuller answer than rules can give. Whether your particular combination of symptoms warrants a call. Whether the article your sister forwarded actually applies to your situation. Whether the medication your GP prescribed has the interaction you are worried about. Those are good questions, and the right answer to them is sometimes a cautiously prompted, well-grounded language model. We would rather build that path correctly, with an opt-in toggle and a sanitized payload, than refuse to build it and watch you ask the same question to a chat window that does not know your pregnancy week.

The reason we wrote this post is the same reason post one in this series existed. The interesting questions about pregnancy apps in 2026 are not about cute features. They are about the boundaries: between local and cloud, between pattern detection and chat, between what an app actually does and what its marketing copy says it does. The honest answer for myCocoon is the one above. Cloud is opt-in. The payload is documented. The default is the quiet one. If you turn it on, you should be able to point at a single screen and know what you have agreed to.

That, more than any one feature, is what we mean by AI in this app.

§