Pilot findings.
Connect Wellness is community-governed AI for health communication. It adapts trusted health information into the multi-dialectal Spanish, English and intermixed language families in the Creighton corridor of Phoenix use every day, checks it with medical librarians, clinicians and residents, and sends it by SMS text message. The community that speaks the language decides what the system may say and when it must stop.
Connect Wellness es inteligencia artificial para la comunicación en salud, gobernada por la comunidad. Adapta información de salud confiable al español de distintos dialectos, al inglés y a la mezcla de ambos que usan a diario las familias del corredor de Creighton, en Phoenix, la verifica con bibliotecarios médicos, profesionales clínicos y residentes, y la envía por mensaje de texto. La comunidad que habla el idioma decide qué puede decir el sistema y cuándo debe detenerse.
¿Prefieres hablar en español? Escríbenos.
The people who speak the language should govern it.
Community data sovereignty, in the spirit of the CARE Principles and the Data is Soil playbook: the people whose words become data decide what is kept close, shared or left alone.
Not yet, and the problem is language.
Families in the Creighton corridor trust the text line they use for food boxes and SNAP. Of 7,976 inbound messages we analyzed, 20 asked about managing diet for a diagnosed condition. The food reaches families; the knowledge to manage a diagnosis has not.
In an April–May 2026 pilot, two community annotators coded 384 messages from 32 patients for health barriers, and nine large language models read the same messages. The annotators agreed with each other far more often than any model agreed with them. Most models scored under 0.5 macro-F1, a measure that weights rare barriers as heavily as common ones.
People imply the rare barriers more often than they state them.
A model that always says "no barrier" gets these 83 right and finds none of these 17.
That is why the chart at the top of this page scores models by macro-F1.
Clinical screening questionnaires ask what institutions decided to ask. When a family stops trusting the system, no form field records it, and the family drops out of care.
Agencies under load are adopting AI navigation tools that often miss the meaning behind people's words. A tool that misreads the people it serves widens the gap it was bought to close. AI in community health should make community health workers more capable and never replace them. Authority over the tools belongs to the people who do the work and live in the neighborhood.
Construction Grammar describes a language as learned pairings of form and meaning, from single words to whole phrase patterns.
Read word by word, the Spanish phrase "Piden muchos requisitos" is "they ask", "many", "requirements", which adds up to a logistics complaint. Read as one phrase with the community library, it means "this system is filtering me out", an eligibility barrier.
We fine-tuned open models on this community's own messages, and none performed acceptably. A fine-tuned model spreads what it learns across millions of numbers no resident can read or fix. A construction is a written record a person can read and correct, and recent research has shown such grammars can be learned from annotated text at scale and stay readable.
So we keep meaning outside the model, in the community phrase-meaning library. The machine proposes entries, the people who speak the language decide which are right, and the library carries over to each new model generation. A standard retrieval-augmented system matches a message's words to documents; the language layer first decides what the message means here, then retrieves, shapes the prompt or routes to a person.
Four library entries from the pilot, adjudicated by community annotators. Council approval is designed.
We test tone, framing and timing with residents. After an exchange, a check-in is designed to ask whether it helped (helped, somewhat or no), and the answers will go to the library and the council.
A measurable condition, counted with a screening question.
A social determinant of health, recorded in the chart.
A matter of dignity, and a reason people hesitate to ask.
All three are true, and none stands in for another. The library keeps facts, phrases, lived experience and community values as separate kinds of record, so one resident's voice is never published as the community's consensus.
Public reference health information is typically written at or above a ninth-grade reading level and built for a web browser. The community's own messages show the reading level, vocabulary and length people use, and six steps bring a trusted passage to that level.
Medical librarians, researchers or doctors curate the source from a trusted public reference and record where it came from.
The passage is atomized into units that each fit one text message and keep a link to the source.
The community shapes it to its own reading level, vocabulary and length, with framing from the phrase-meaning library.
Four gates check it, in sequence: source quality, community communication review, nutritional accuracy, and clinical safety when the topic calls for it; a unit that misses community review goes back for re-adaptation.
It goes out where people already are, by text in Spanish or English, in the channel families use for food.
Creighton Community Foundation
Text message
Calabacitas, nopales y chayote casi no suben el azucar. Cook them with onion and tomato and fill half your plate. Rice or tortilla goes on the side.
Las tortillas de maiz son grano entero y tienen mas fibra que las de harina. Corn tortillas con frijoles y verdura make a good, filling meal.
Living curation keeps it current: we will review sources quarterly on a library platform and send changed units back through the gates.
Source changes: back to step 2
The system never gives dosing, a diagnosis or individualized medical advice. Medication and symptom questions go to a provider.
Pending funding, building with the ASU Library Health Sciences Division, a member of the Network of the National Library of Medicine, Region 5.
The community is designed to decide. Connect Wellness is built as community-based participatory research (CBPR) and youth participatory action research (YPAR), with residents, youth and community health workers as co-researchers.
A paid, community-majority council
Steps with an orange edge are council decisions.
Meaning what a phrase means
Residents and community health workers will teach the library in ordinary conversation, with stories, examples and corrections. Contributors will be attributed for their entries and paid for the work.
Both pilot annotators are community members, and one is also research-trained. On hard cases, the annotator without research training deferred to the one with it. The codebook did not account for that power dynamic; a council where community members hold the votes is this design's correction.
Two annotators code every message, and hard cases are designed to go into a registry. We report where the community's categories differ from standard clinical screeners.
Through the same text line, in Spanish and English.
Illustration of the design. Not a pilot conversation.
"Around here, asking for help can feel like admitting you failed your family."
Is this your own experience, something you see around you, or something handed down?
About 26,000 messages a year come through our community text line. Read together, they say what households need, what is changing, and where the systems around them fall short.
Frameworks that help health systems understand community obstacles and needs are informed differently, and become better. Issues driven by economic or socio-political forces surface sooner: the loss of SNAP benefits from new enrollment processing surfaces now, rather than after 200,000 people have lost food support.
Then the loop starts again with new household messages.
When the community speaks, institutions answer on the record.
This section describes work in progress; none of it is a finished system yet.
We are building the community phrase-meaning library and the teaching mechanism first. The knowledge graph sits underneath them and connects each entry to the people, places, institutions and services it refers to.
Two goals shape the design from the start. Another community should be able to run its own library, under its own permissions. And any AI application, in education, health care, human services or civic life, should be able to consult the language layer through an interface that works with any model.
Giving an AI system the community's interpretation does not guarantee that it will follow it. That is why the four gates and the council's go and stop authority exist.
When a message carries a barrier, its context should reach the clinical record as a referral a clinician can act on, and we are building the instrument that scores whether it did.
A community organization that already has the text platform, the consented messages and the relationships, with partners who hold the work to a research standard.
Creighton Community Foundation, in the corridor since 2013, and a team of community health workers who live in the neighborhood.
Arizona State University College of Health Solutions; ASU Linguistics, Construction Grammar; ASU nutritional epidemiology.
ASU Library Health Sciences Division, a member of the Network of the National Library of Medicine, Region 5.
Banner University Medical Center; University of Arizona College of Medicine Phoenix.
UCLA, and a national NIH-supported AI research consortium in which CCF is a community partner.
MessageEverywhere, the MIT-licensed messaging platform CCF built with an open developer community.
An HRSA community health worker apprenticeship partnership.
Funders, health systems and governments can bring a population, a clinic or a program. Technologists can build on the MIT-licensed platform. Community members can ask about a council seat. Email the team and a person replies.
Can a community author the semantic infrastructure through which AI understands it?
Community-governed AI for health communication adapts and delivers health information in the language a community uses, with the community deciding what the system may say and when it must stop. At Connect Wellness that authority is designed to sit with a paid, community-majority council of residents and community health workers.
No. The system helps community health workers spot needs sooner and get people to them faster, and the workers are designed to govern the tools. A message that carries fear, confusion or withdrawn trust goes to a person.
Construction Grammar treats language as pairs of form and meaning. A community phrase-meaning library records how this community expresses need, fear and trust in multi-dialectal Spanish, English and intermixed language, so a system can recognize the meaning behind the words.
We tried. Fine-tuned open models did not perform acceptably on this community's language, and nobody in the neighborhood can inspect or correct a fine-tuned model. Residents can audit and extend a library that sits outside the model, and it carries over when models change.
Medical librarians, researchers or doctors curate the source. We cut the passage into a text-sized unit that keeps its link to the source, reword it in community language, and check it at four gates: source quality, community communication, nutritional accuracy and clinical safety.
The community council is designed to decide. Each week it will review what the system caught and missed, approve library entries, and decide when a model may message families and when it must stop.
Messages are collected with participant consent during routine operations; the corpus is de-identified; research use runs under an IRB of record (ASU). Names, phone numbers and other identifying details never appear in published examples.
Email the team at [email protected]. Funders and health systems can bring a population or a program. Technologists can build on the MIT-licensed platform. Donors can fund community council seats directly.
