The paper is soft from folding. A grid of boxes and dotted lines that has traveled in a sari pleat or a shirt pocket across monsoon puddles and up four flights of unlit stairs. It’s a TB contact-tracing form, used by an ASHA worker in Mankhurd, Mumbai. Spaces for names, ages, symptoms, sputum test dates. And margins. In those margins, the worker has written things the form did not ask for: husband left, no income this month, scared of hospital, will go if sister accompanies. These aren’t data points. They’re the difference between a contact who gets traced and a contact who gets lost.
I’ve spent the last decade watching community health workers navigate the gap between what a protocol imagines and what a doorway reveals. Scripts proliferate in that gap. Some are paper. Some are digital checklists on government-issued phones. Some are oral—a sequence of questions a worker has learned to ask in a particular order because she’s found, over hundreds of visits, that this order opens a conversation rather than shutting it down. These scripts are technologies of coordination. They’re also technologies of power. And they’re already here, already being generated, long before anyone mentions an AI screenplay tool.
I want to think about what these existing scripts do, and what they undo, before we rush to replace them with something that claims to be smarter. The logic of script generation—the idea that a well-designed template can guide a human interaction toward a desired outcome—is not new to public health. It’s embedded in every job aid, every counseling algorithm, every decision-support flowchart that tells a frontline worker what to ask and when to refer. The question isn’t whether we should use scripts. It’s who writes them, who gets to deviate from them, and whose knowledge they render invisible.
The Paper Script as Infrastructure
In Mumbai’s informal settlements, ASHA workers carry a small booklet for maternal health visits. Each page corresponds to a trimester, with columns for blood pressure, weight, fetal heart rate, danger signs. The booklet is a product of the National Health Mission, designed in Delhi, refined through state-level trainings, printed in bulk. It is, in a very real sense, a script: it tells the worker what to ask, in what order, and what to record. It also tells her what not to ask. No column for whether the woman has eaten today. No column for whether her husband has threatened to throw her out if she delivers a girl. No column for the smell of kerosene in a windowless room where a pregnant woman sleeps.
I’ve sat with ASHA workers as they fill these booklets. They do something remarkable: they follow the script and they exceed it. They ask the required questions, record the required numbers, and then they stay. They ask about the last meal. They notice the kerosene. They write it in the margins, or they hold it in memory, or they pass it to the anganwadi worker who handles nutrition supplements. This improvisation isn’t a failure of the script. It’s the script’s necessary supplement—the human labor that makes a standardized protocol work in a non-standardized world.
But here’s what worries me: when we digitize these scripts, when we turn them into phone-based job aids with mandatory fields and dropdown menus, the margins disappear. The ASHA worker in Mankhurd now carries a phone with an application that prompts her through TB contact tracing. She must complete each screen before advancing. The application timestamps her entries. It geotags her visits. It generates performance dashboards for her supervisor. It also eliminates the space where she used to write scared of hospital. That information now lives only in her head, if it lives at all. It isn’t aggregated. It isn’t analyzed. It doesn’t shape policy. It’s lost to the system, even as it remains crucial to the care.
What the Script Silences
In Bengaluru, I worked with a team of urban health navigators—community members trained to support tuberculosis patients through treatment. They used a counseling script developed by a respected research institution. Evidence-based, culturally adapted, pilot-tested. It covered medication adherence, side-effect management, nutrition, and stigma reduction. By any technical standard, excellent.
But the navigators kept deviating from it. Not randomly. Systematically. They learned that starting with nutrition—asking whether the patient had eaten that day—opened a conversation that the script’s opening question about medication adherence did not. They learned that patients who were hungry weren’t ready to talk about side effects. They learned that asking about stigma before establishing trust produced polite, useless answers. They rewrote the script through practice, and they did it without permission.
When we interviewed them about their deviations, they described a kind of embodied knowledge. “You can feel when a house is ready for certain questions,” one navigator said. “The way they offer you water, or don’t. The way they look at the door.” This knowledge isn’t in the script. It can’t be in the script, because it’s relational and contextual and acquired through the body of a worker who has been in hundreds of homes. It’s the kind of knowledge that automated script generation—which learns from data, not from doorways—cannot access.
This isn’t an argument against structure. The navigators didn’t abandon the script; they adapted it. They used it as a scaffold, not a cage. But the adaptation depended on their authority to deviate, and that authority is precisely what many digital health platforms are designed to eliminate. When every deviation is flagged as non-compliance, when every skipped question generates an alert, the worker’s knowledge becomes a liability rather than a resource.
The Generator in the Room
There’s a seductive logic to automated script generation. It promises consistency. It promises to scale expertise. It promises to reduce the burden on overworked community health workers by giving them exactly the right thing to say at exactly the right moment. These aren’t trivial promises. In a system where ASHA workers are expected to manage multiple health programs with minimal training, a well-designed decision-support tool can genuinely improve care.
But the promise depends on a particular model of what a health interaction is. It imagines the interaction as a series of decision points: if the patient says X, the worker should respond with Y. It imagines that the relevant information is what can be captured in discrete data fields. It imagines that the worker’s primary task is information extraction and protocol execution. It imagines, in other words, a world where the generator is not in the room.
The generator is not in the room. It doesn’t smell the kerosene. It doesn’t see the woman’s face when her husband enters. It doesn’t feel the hesitation before a question about alcohol use. It doesn’t know that this particular patient’s sister is a nurse, or that this particular lane floods in June, or that this particular family lost a child to dengue three years ago and is terrified of fevers. All of this knowledge shapes what a skilled community health worker says and doesn’t say, asks and doesn’t ask, records and doesn’t record. It’s the knowledge that makes the script work.
When we remove the generator from the room—when we centralize script production in an algorithm trained on aggregated data—we’re not just changing the technology. We’re changing the epistemology. We’re deciding that the knowledge that matters is the knowledge that can be extracted from datasets, not the knowledge that accumulates in bodies that have walked through thousands of doorways. We’re deciding that consistency across contexts is more valuable than responsiveness to context. We’re deciding, in effect, that the margins of the form should remain blank.
The Labor of Listening
There’s a deeper problem, and it has to do with what we count as work. When an ASHA worker fills a TB contact-tracing form, the system recognizes her labor: she has completed a task, generated a data point, contributed to a performance metric. When she sits with a woman who is scared of the hospital and talks through that fear for twenty minutes, the system recognizes nothing. That time isn’t billable, not measurable, not visible on any dashboard. From the perspective of the script, it’s a deviation.
But that deviation is often the intervention. I’ve tracked TB treatment outcomes in Mumbai long enough to know that the patients who complete treatment aren’t necessarily the ones who received the most technically correct counseling. They’re the ones who had a health worker who noticed when they stopped answering the phone, who knew which neighbor to ask, who showed up at the door with a food packet and no judgment. These actions aren’t in any script. They’re the product of a relationship, and relationships require time that no algorithm can optimize.
The push toward automated scripts is partly a response to scarcity: not enough health workers, not enough time, not enough training. Automation promises to stretch scarce resources. But it also risks redefining the work in a way that makes the most important parts of it invisible. If we measure success by script compliance, we’ll reward workers who follow the protocol and penalize workers who follow the patient. We’ll generate data that shows high fidelity and low impact, and we won’t understand why.
What the Bengaluru Navigators Taught Me About Scripts That Work
The urban health navigators in Bengaluru eventually developed their own script. It wasn’t written down. It was a shared oral protocol, transmitted through shadowing and debriefing and the kind of collective sense-making that happens when workers sit together after a day of visits and talk about what worked. The protocol had a structure: start with food, move to symptoms, address stigma only after trust is established, always ask about the household before leaving. But it also had a principle: the structure is a guide, not a rule. Deviate when the context demands it. Trust your reading of the room.
This protocol was more effective than the official script, by measures that mattered: treatment completion rates, patient-reported trust, early detection of side effects. But it was also more demanding. It required workers to be skilled listeners, skilled observers, skilled improvisers. It required them to carry the cognitive load of constant contextual judgment. It required them to be generators, not just executors.
This is the tradeoff that automated script generation obscures. Automation reduces cognitive load—a genuine benefit for overworked health workers. But it does so by shifting authority away from the worker and toward the algorithm. It makes the worker a relay rather than a decision-maker. And in doing so, it may reduce the very capacities that make community health workers effective: their ability to read a room, to build trust, to adapt in real time to the specific human beings in front of them.
The question isn’t whether we should use scripts. We already use them, and we’ll continue to use them. The question is what kind of scripts we design, who designs them, and what space they leave for the knowledge that can’t be specified in advance. A script that leaves room for the margins—that treats deviation as information rather than error—is a very different technology from a script that enforces compliance. The difference isn’t in the code. It’s in the power relations the code encodes.
Who Was in the Room?
Every script has authors. The maternal health booklet was authored by obstetricians and public health administrators in Delhi. The TB counseling script was authored by researchers and behavior-change specialists. The oral protocol of the Bengaluru navigators was authored by the navigators themselves, drawing on their experience and their knowledge of their communities. These different authorships produce different kinds of scripts, and different distributions of authority.
When we talk about automated script generation, we need to ask: who is training the algorithm? What data is it learning from? Whose interactions are represented in that data, and whose are missing? The ASHA worker in Mankhurd who writes scared of hospital in the margin is generating knowledge that no dataset captures. If we train an algorithm on the formal data—the completed fields, the timestamped entries, the performance metrics—we’ll produce scripts that optimize for what the system already values. We won’t produce scripts that optimize for what the margins contain.
This isn’t a technical limitation. It’s a structural one. The knowledge that matters most for effective community health work is often the knowledge that is least legible to formal systems. It’s relational, contextual, embodied. It’s passed through apprenticeship rather than training modules. It’s stored in memory rather than databases. Any script-generation technology that doesn’t grapple with this epistemology will produce scripts that are technically sophisticated and practically hollow.
The Authors Guild, in its AI Best Practices for Authors, emphasizes the importance of human creators maintaining control over their work and being compensated for its use. The same principle applies here, though the “authors” aren’t novelists. They’re community health workers whose improvisations, whose marginal notes, whose relational labor constitutes a form of authorship that current AI frameworks don’t recognize. If we’re going to automate script generation, we need to ask: whose scripts are we learning from, and are they being compensated, credited, or even asked?
The Maintenance Work That Never Makes Headlines
There’s a final thing I want to say, and it’s about maintenance. The paper scripts I’ve described—the booklets, the forms, the oral protocols—require maintenance. They require workers who update them, adapt them, argue about them, and teach them to new colleagues. This maintenance work is invisible to the health system. It isn’t funded, not supervised, not evaluated. It happens in the interstices: during tea breaks, on shared auto rides home, in WhatsApp groups that no program officer monitors.
When we automate script generation, we’re not eliminating maintenance. We’re shifting it. Someone will need to update the algorithm. Someone will need to monitor its outputs for errors and biases. Someone will need to decide when a deviation is a feature and when it’s a bug. The question is who that someone will be. If it’s a data scientist in a Bangalore office who has never sat in a Mankhurd doorway, the maintenance will optimize for different things than if it’s the ASHA worker herself. The script will drift toward the concerns of the people who maintain it.
This is why the question of who is in the room matters so much. It matters at the moment of design, but it matters even more at the moment of maintenance, because maintenance is where scripts actually become what they are. A script maintained by the people who use it will evolve toward context. A script maintained by the people who fund it will evolve toward metrics. Different evolutionary pressures, different creatures.
Leaving the Margins Open
I’m not arguing against technology. I’ve spent too many years watching community health workers struggle with paper forms that get lost, with data that never gets analyzed, with protocols designed for populations that don’t resemble the populations they serve. There are real problems that better tools could address. But I’m arguing against a particular kind of technological solutionism that treats human judgment as a problem to be engineered away rather than a resource to be supported.
The best scripts I’ve seen in community health aren’t the ones that specify every action. They’re the ones that provide structure while preserving discretion. They’re the ones that treat the worker as a generator, not an executor. They’re the ones that leave the margins open, literally or figuratively, for the knowledge that emerges in the encounter itself. Designing such scripts is harder than designing compliance-enforcing algorithms. It requires humility about what can be specified in advance. It requires trust in the people who do the work. It requires a willingness to measure success by outcomes that are messier than script fidelity.
Creative writing pedagogy, as the Purdue OWL resources on creative writing suggest, often emphasizes the importance of understanding conventions in order to deviate from them effectively. The same principle applies to community health scripts. The goal isn’t to eliminate structure but to cultivate the judgment to know when structure serves the encounter and when it doesn’t. That judgment can’t be automated. It can only be developed, through practice, through mentorship, through the slow accumulation of doorway experiences that no dataset contains.
The ASHA worker in Mankhurd still carries her paper form. She still writes in the margins. The phone in her pocket prompts her through the required fields, and she completes them, and then she does the work that the phone doesn’t know about. She’s already navigating two scripts: the formal one that the system recognizes, and the informal one that actually guides her practice. Any new technology we introduce will enter this ecology. It will either make space for the second script or it will try to eliminate it. That choice isn’t technical. It’s political. It’s about who we trust to know what care requires, and whose knowledge we’re willing to make visible.
The scripts we already carry are imperfect. They’re incomplete. They’re maintained by people who are underpaid and overworked and rarely consulted about the tools they use. But they contain something that no automated generator can replicate: the residue of thousands of encounters between a worker and a patient, a doorway and a body, a question and a silence. Before we replace them, we should understand them. Before we automate them, we should ask what they already know.