Artificial Intelligence in Perinatal Education: Opportunities, Risks, and Ethical Considerations

By Birth & Baby University · Last updated: October 2026 · 7 min read

Pregnant and postpartum people are already asking AI chatbots questions, and educators are being offered tools that promise to draft handouts, answer client questions and even act as a “digital doula.” The technology is moving faster than the evidence. A small but growing group of studies tests how accurate and readable chatbot answers are, and a few papers lay out what safe use would require. This article reviews what those studies show, where the risks lie, which ethical questions apply to doulas and childbirth educators, and how to use these tools, if at all, without compromising clients or scope.

Key takeaways

Quick answer: in studies that put standardized questions to chatbots, accuracy on knowledge tests can be high, but answers are often written above the recommended reading level, give little source information and can fall short of clinical safety standards. No source reviewed shows that AI tools improve outcomes for perinatal families. The sensible position is to treat them as drafting aids that need human review, not as sources of advice.

• In a 2026 comparison of six chatbots on 200 multiple-choice questions about postpartum depression, the best model scored 97.50 percent, but transparency and source attribution scored low and all models exceeded the recommended sixth-grade reading level.
• In a second 2026 study of five chatbots answering 27 perinatal depression questions, most answers were empathetic and informative but fell short of fully addressing clinical safety standards, and all exceeded the sixth-grade level.
• A 2026 proposal for “digital doulas” argues for AI as an adjunct inside a stepped-care model with human oversight and warns of the illusion of empathy, crisis recognition failures, bias, privacy vulnerabilities and unsafe deployment.
• A systematic review of 15 chatbot studies reported high acceptability and no serious harms, but its conclusions are strongly worded and the studies are varied, so they should be read cautiously.

What studies of chatbot answers show

Yang and colleagues (2026) tested six chatbots with 200 standardized multiple-choice questions on postpartum depression. ChatGPT-5 reached the highest accuracy (97.50 percent). In reliability and readability tests on 20 public-education questions across four models, ChatGPT-5 scored highest on content-quality instruments, but scores on a benchmark for transparency, source attribution, currency and disclosure were low across all models. Every model produced output above the recommended sixth-grade reading level. The authors state that the study did not evaluate clinical safety, patient comprehension, user behavior or real-world effectiveness, and conclude that chatbots should not be used as autonomous sources of postpartum mental health information.

Huang and colleagues (2026) put 27 frequently asked perinatal depression questions to five chatbots and had two obstetricians rate the answers with validated tools. Models differed on quality measures, with different models scoring highest on different instruments. All exceeded the National Institutes of Health recommended sixth-grade reading level, with the automated readability index averaging between 13.49 and 15.81. Most models produced empathetic and informative content but fell short in fully addressing clinical safety standards. These findings describe particular models in a particular period, and chatbots change often, so the results may not hold for the version a family uses today.

The reviewed studies share limits. They test text answers, not what people do with them, and they focus on one topic, perinatal depression. They do not tell us about breastfeeding, newborn care, labor decisions or other subjects that doulas and educators cover.

What the broader evidence base says

Rasoli and colleagues (2026) systematically reviewed 15 studies of generative conversational agents for pregnant people or their partners. They report mean acceptability scores of 85 to 94 percent, retention of 62 to 78 percent over six months, and no serious harms. The conclusion is enthusiastic, calling these tools a breakthrough and supporting rapid integration. Readers should be cautious about that framing. The included studies span different tools, countries and designs, and acceptability and retention are not clinical outcomes. A review that finds no reported harms is not proof of safety, since its abstract does not say the studies were designed to detect harms.

Yang and Lin (2026) take a more conditional view. They argue that digital doulas should be AI-enabled relational interfaces within a stepped-care model, not substitutes for clinicians or human doulas, with four proposed functions: companionship, symptom interpretation, navigation and sentinel monitoring. They identify significant risks, including the illusion of empathy, crisis recognition failures, algorithmic bias against vulnerable populations, data privacy vulnerabilities and unsafe deployment, and recommend human oversight, clear escalation protocols with medico-legal accountability, workflow integration and equity-centered evaluation. This is a proposal, not a test.

Ethical considerations

The World Health Organization’s guidance on ethics and governance of artificial intelligence for health says AI technologies hold promise but must put ethics and human rights at the heart of their design, deployment and use. It identifies ethical challenges and risks and sets out six consensus principles. This article draws only on the publication overview and does not summarize those principles. Applied to perinatal education, the questions that follow from the studies above include:

• Accuracy and currency. Is the information current and consistent with professional guidance, and can the user see where it came from?
• Readability and access. Answers above the sixth-grade level may leave behind people with lower health literacy, the same families who may most need clear information.
• Safety and escalation. What happens when a user describes a crisis, thoughts of self-harm or an urgent symptom?
• Privacy. What information is collected, who can see it and how long is it kept?
• Bias and equity. Does the tool work equally well across languages, cultures and communities?
• Transparency. Do families know when they are talking to software, and when content was produced or drafted with AI?

What this means for doulas and educators

• Treat AI output as a draft. Check every factual claim against a primary source such as a professional guideline or government agency before using it with clients.
• Do not use AI for symptom advice. Doulas do not diagnose, and chatbots have not been shown to be safe for that purpose. Direct clients to their care team or emergency services.
• Protect client information. Do not enter names, health details or other identifying information into a tool unless you understand and accept how it stores and uses data.
• Check readability. If you use AI to simplify text, test the result for plain language, and ask a colleague or client to review it.
• Be transparent. Tell clients if a handout or message was drafted with AI and reviewed by you.
• Do not outsource the relationship. The studies caution about the illusion of empathy. Human presence is the core of doula work.

Putting it into practice

1. Write a one-page AI use policy for your practice covering what you will and will not use it for, how you check output and how you protect data.
2. Keep a list of the primary sources you rely on, so reviewing AI-drafted text is quick.
3. When clients tell you they use chatbots, ask what they learned and offer to go through it together, without judgment.
4. Give clients a short list of trusted sources and urgent warning signs from a reputable agency, and tell them what to do in an emergency.
5. Revisit your policy every few months, since tools and evidence change.

Limits and uncertainty

Most sources here are 2026 papers on specific chatbots, tested on one topic, and the tools they tested will have changed by the time you read this. One systematic review draws strong conclusions from varied studies. The digital doula paper is a proposal. No source tests whether AI use by doulas or educators improves outcomes or whether it causes harm. The WHO source was read at overview level only.

The bottom line

AI tools can produce fluent, often accurate-sounding answers, but the research so far shows gaps in transparency, readability and clinical safety, and it does not show benefits for families. For doulas and educators, the practical response is human review, plain language, privacy care and honesty with clients about how the tools are used.

Birth & Baby University’s For Doulas page outlines its online, self-paced birth and postpartum doula training, recertification and mentorship, for doulas who want grounded, human-led skills.

This article is for professional education. It is not medical, legal, financial or tax advice. Check the standards of your certifying organization and the rules where you work.

Sources and further reading

• Yang M, Liu H, Lin S, et al. Evaluating the accuracy, reliability, and readability of AI chatbots in delivering postpartum depression information. Frontiers in Psychiatry (2026). Read the source
• Huang J, Yu H, Chen J, et al. Can large language models be trusted? Reliability and readability of responses to perinatal depression FAQs. Frontiers in Public Health (2026). Read the source
• Yang H, Lin DM. Conversational AI for perinatal mental health: promise, limits, and a human-AI stepped-care framework. Frontiers in Psychiatry (2026). Read the source
• Rasoli R, Ebrahimisadrabadi F, Khedri Z, Sohrabei S. Designing conversational intelligence: effect of large language models (GPT-driven) platforms for precision maternal and newborn health engagement: a systematic review. Oxford Open Digital Health (2026). Read the source
• World Health Organization. Ethics and governance of artificial intelligence for health (guidance, 2021). Read the source

More for professionals: the Professionals Blog.