Medha Thapliyal

Sheet 03 Writing Revised October 2026

Sheet 03/04. 17 February 2026. 3 minute read.

Code-mixed and context-less

Real maternal health queries are short, mix languages, and leave out the symptom. The safety problem in chatbots for mothers is linguistic and social before it is technical.

03/04 · Essay

The largest maternal messaging program in the world is Indian. A voice-call service built by a Mumbai non-profit has reached tens of millions of women and children with timed messages through pregnancy and infancy. It also loses nearly a quarter of its listeners to dropout, and the most interesting AI work around it has been unglamorous: a model that predicts when a woman is likely to answer the phone, which improved pickup by about twelve percent in a pilot district. No chatbot. No diagnosis. Just a better guess about whether to call at ten in the morning or six in the evening.

I think about that result whenever a new maternal health chatbot is announced, because it shows where the leverage actually is. The next generation of tools is more ambitious. Messaging assistants in Hindi, Assamese, and English, built on language models, that answer a pregnant woman’s questions directly. A preprint from earlier this year analysed real queries to one of these systems and found three things that anyone who has sat in an antenatal clinic would recognise.

What the queries look like

They are short. A few words, often a single noun. They are code-mixed, moving between Hindi and English and a regional language inside one sentence, with spellings that follow speech rather than any dictionary. And they are missing the context a clinician would need: a woman asks about pain without saying where, or about bleeding without saying how much or when.

This is not a failure of the users. It is how people talk to a phone when they are worried, when data is expensive, when they are typing with one hand, and when the question they most need answered is one they are embarrassed to ask directly. A model that was trained mostly on English text written by people with time to write full sentences is being asked to triage from fragments.

Why this is a safety problem first

The technical response is to improve the model: more Hindi data, better transliteration handling, prompts that ask clarifying questions. All reasonable. But the safety problem is upstream of the model. A system that answers “pet mein dard” with general reassurance about pregnancy discomfort has made a triage decision on no information, and the person who bears the consequence of that decision is a woman with no other source of advice at that hour.

In the behaviour change communication work I contributed to, the single most important design decision was never the content. It was the pathway: what the message tells a woman to do, who she can reach, and whether that person will be there. A chatbot that cannot route a worrying fragment to a human who can call back is a leaflet that talks.

What I would want from a maternal chatbot

I would want it to be bad at reassurance and good at escalation. Every ambiguous query about pain, bleeding, fever, or reduced movement should produce a human contact, not an answer. I would want it evaluated in the language people actually use, including the misspelt and the mixed, with error review by midwives who work in the districts it serves. I would want a published account of what it does when it does not understand. And I would want its success measured not in messages answered but in the number of women who reached a facility in time because of it.

The call-timing model is still my favourite piece of AI in this space. It did one small thing, measured it, and made an existing program better. That is the bar.

Sources

03 · Adjoining

Keep reading.

  1. Aug 2026 The red alert and the ASHA A clinical copilot in Nairobi cut diagnostic errors most when it interrupted. What would an interruption mean for a community health worker with no clinical backstop?
  2. Jul 2026 Novices with a probe One day of training produced sonographer-level gestational age estimates in Lusaka. The hard question is what happens to the referral chain after the machine says high risk.
  3. Jun 2026 Silicon mothers Language models can produce plausible answers about pregnancy in Assam. Plausibility is not validity, and formative research cannot be outsourced to a model that has never missed an antenatal visit.