By Hawra Rabaan
May 27, 2026
My research on domestic violence, help-seeking, and healing within Muslim communities (in the US and beyond) shows that women’s agency is not just a matter of individual choice or resistance; it reflects how women act within constraints shaped by social, economic, and political systems. Autonomy, then, is about whether those constraints allow for meaningful choice.
For the women I’ve interviewed, healing from violence rarely unfolded as a single decision or a single event. It was negotiated. Relational. Layered across family, faith, law, and community. As AI-powered applications aimed at safety decision making for partner violence became an option in the healing process, promises of personalization, earlier detection, and efficiency at scale raise questions about the assumptions being built into these systems. How are the systems designed, what data are they trained on, and whose voices shape them? When the complex lives of non-dominant communities are flattened into datasets, who disappears?
When Harm Is Hard to Name
In my research, women often described their experiences of economic control, educational deprivation, restriction from work, and psychological coercion — but they did not initially name it as violence. Across US and non-Western contexts, women’s understandings of domestic violence were fluid rather than fixed. They often recognized harm gradually, through comparison, reinterpretation, and exposure to alternative narratives. Practices framed as “normal” were later understood as coercive. This variability was not confusion; it reflected how domestic violence is embedded within family structures, cultural norms, and institutional systems.
Many emerging tools aimed at assessing family and domestic violence draw on administrative data, including police incident reports and criminal history, electronic health record–based screening programs and alerts, and child welfare–level screening scores. While these systems capture incidents, they miss the slow cognitive, relational, and sometimes theological labor of recognizing harm. Recognizing harm is not a discrete step that precedes decision-making; it unfolds by ongoing risk assessment. Women come to recognize abuse gradually—unlearning normalized behaviors and confronting entrenched norms, continuously weighing social stigma, economic dependence, and legal uncertainty as they interpret their experiences. As a result, recognition is partial and contingent. This creates a fundamental problem for AI systems that assume harm is already legible and separable from risk. Models built on observable indicators overlook the fact that the recognition of harm is shaped by lived calculations of risk. In doing so, these systems transform a constrained version of harm into something that appears objective.
AI systems learn from datasets that are built on institutional definitions of violence, prioritizing what is measurable and recorded, while overlooking less legible forms of harm. When abuse does not conform to predefined categories (such as visible injury, or a documented report), it disappears within risk assessment systems, rendering lived experiences only partially visible or entirely unaccounted for. For example, although the US Department of Justice’s definition of abuse encompasses multiple forms of domestic violence, police reporting systems operationalize these definitions through legally actionable incidents. In practice, this narrows what counts as abuse to what can be acted on legally or clinically, excluding forms that are relational, cumulative, or culturally mediated — such as educational deprivation, restriction from work, religious manipulation, or extended family-based coercion. If definitions of abuse are socially shaped and dynamically negotiated, AI systems built on universal, fixed categories will not merely miss complexity, they will reproduce institutional blind spots at scale.
When Patience Is Misread as Passivity
Even after women recognized their abuse, their departure was not always immediate or preferred. Some women negotiated. Calculated. Waited. One woman saved money quietly for years; she navigated domestic expectations while subtly expanding her autonomy. From the outside, this might have looked like endurance or passivity. From within, it was a strategy. Calculated patience was not the absence of agency. It was agency under constraint, a careful assessment of when resistance would be survivable.
Health and legal systems often register survivors’ agency primarily through visible, documentable acts (for example, filing restraining orders, entering shelters, or engaging in formal treatment). Flattened datasets cannot see strategic survival; they reduce complex decision-making to binary categories such as whether a person leaves a situation or stays in it. When that logic is embedded into AI triage systems, strategic patience can be misclassified as denial or non-compliance.
When Marriage Exists in Two Systems
Another layer complicates this picture: for many Muslim women, marriage does not exist in a single institutional domain. It is often simultaneously civil and religious. Divorce, too, may unfold unevenly across those systems. A woman may be legally divorced in a US court but remain religiously married in her community’s eyes. Conversely, a marriage may be religiously solemnized but never legally registered.
Abusers can exploit this dual structure strategically — divorcing in one system but not the other, threatening to withhold religious dissolution, leveraging immigration status as coercion.
Health systems, however, typically record marital status as a binary field: married, divorced, single. Flattened datasets assume marriage exists in one institutional framework, dismissing the lived reality of navigating overlapping legal and religious authorities, and failing to recognize how this split shapes mental health, financial precarity, and help-seeking decisions.
When AI systems rely solely on civil records, they misread vulnerability. A woman coded as “divorced” may remain socially constrained. A woman coded as “married” may lack legal protections. The data appears coherent. The lived reality is not.
When Healing Requires an Ecosystem
In my research, domestic violence was never a two-person story. It was an ecosystem.
Male allies, including imams and influential community members, sometimes intervened when a woman’s voice was dismissed. Older women also played complex roles. Some grandmothers modeled resilience and encouraged girls’ education. Others mediated disputes. In some cases, older women normalized endurance and prioritized family reputation over individual safety.
Faith leaders occupied a similar paradox. When trained and accountable, they reframed sacred teachings in ways that affirmed justice. When untrained or isolated they sometimes dismissed, minimized, or misunderstood survivors’ experiences, reinforcing harm rather than interrupting it. The advocates and social workers who made the greatest difference were culturally competent, trauma-informed, and deeply connected to networks of support. They understood immigration law, religious nuance, modesty norms, and systemic bias. This relational infrastructure — what I call a healing structure — helped sustain safety, support decision-making, and maintain care over time.
Flattened datasets reduce this ecosystem to individuals, incidents, and risk scores. A woman becomes a case file. Harm becomes a checkbox. Context becomes metadata. The complexity of influence, elders, faith leaders, immigration status, reputational risk, economic dependence, community mediation, collapses into structured fields.
In that translation, relationships become invisible and cases become linear. But healing does not move in straight lines. It unfolds through webs of influence: through the advocate who anticipates backlash, the imam who reframes scripture, the elder whose approval shifts power, the social worker who understands how immigration law can be weaponized. When AI systems are designed for efficiency rather than relationships they may streamline access to services while weakening the very infrastructures that make those services usable. A chatbot can deliver information, but it cannot absorb stigma. A triage model can flag risk, but it cannot redistribute authority within a gendered community. A digital “therapist” can offer coping tools, but it cannot buffer retaliation. If AI in women’s health ignores the ecosystems where harm and healing actually happen, they risk increasing efficiency at the cost of depth, accelerating processes while thinning care.
Dimensionality and the Future of Women’s Health
Countering this tendency toward flattening requires two commitments: to plurality and localization. By plurality, I mean resisting singular, authoritative answers. Within Islamic feminist thought, knowledge is interpretive and contextual; justice is relational; even concepts like marital harmony are debated rather than fixed. A plural AI system would not issue a single directive. Instead, it would surface competing interpretations, acknowledge uncertainty, and allow users to navigate complexity gradually. Drawing from slow design principles, such a system would support users over time instead of pushing immediate decisions. It would acknowledge that recognition, harm reduction, and decisions about staying, negotiating, or leaving are often incremental processes shaped by layered constraints. Rather than optimizing for rapid compliance or instant triage, it would support paced reflection, iterative decision-making, and strategies that minimize harm while expanding viable options within the user’s own context.
By localization, I mean recognizing that women navigate specific legal, cultural, and spiritual terrains. US-based Muslim women confronting violence may face immigration constraints, gendered Islamophobia, faith-based mediation processes, and powerful community norms. An AI system trained on generic datasets will miss these layers. Yet localization itself carries risks: it can slide into essentialism, freezing culture into code. We must also guard against cultural relativism, dismissive universality, and cultural reductionism. The challenge is to design localized systems that remain internally plural, grounded in context while preserving complexity rather than fixing culture into rigid categories. Instead of asking, “What is this user’s risk score?” we might ask, “What ecosystem shapes her health decisions, and how can technology strengthen that ecosystem rather than override it?”
Plurality and localization do more than correct bias. They reconfigure intervention. They shift AI from prediction engines to relational scaffolding, from centralized authority to distributed accountability, from behavioral monitoring to the expansion of autonomy. This matters, because the value of AI lies in how it is embedded within broader relational and institutional structures. If trained on culturally competent and legally localized data, AI systems could better tailor information to users’ specific circumstances—something current systems do poorly. More importantly, AI can function as a connective layer, linking users to relevant resources, interpretations, and networks that are otherwise difficult to access.
In my research, participants often relied on search engines to understand their situations, yet these tools consistently failed them. Legal information was too general to be actionable, and religious guidance was shaped by profit-driven rankings and dominant institutional voices, making patriarchal interpretations most visible while feminist perspectives were buried. Some participants turned to Facebook groups, but resources there were fragmented, hard to find, and not always relevant. For users already navigating uncertainty, stigma, and risk, this landscape was overwhelming, discouraging, and often misleading.
A system trained on plural, community-grounded sources — including Islamic feminist interpretations of abuse across both marital, non-marital, and family contexts — could surface perspectives that are currently marginalized. It could help distinguish between cultural practices and religious principles, offering a more nuanced understanding of one’s situation. For women with limited access to diverse sources of knowledge, this could be transformative.
Poorly designed systems risk reproducing harm at scale. But carefully designed, localized, and plural systems could help counter informational inequalities and make complex decisions more navigable. If we invest in community-embedded AI models, culturally competent training for health actors, hybrid human-AI accompaniment systems, and evaluation metrics that prioritize autonomy and trust, we begin to build a different future — one in which technology strengthens the infrastructures that make survival possible.
Dr. Hawra Rabaan is a design researcher and sociotechnical scholar. Her work is grounded in feminist and social justice approaches to computing and design. Rabaan introduced Islamic feminism as a theoretical and analytical lens in human-computer interaction, and developed the survivor-centered transformative justice framework to address domestic violence by centering survivor autonomy, fostering community accountability, and advancing institutional transformation. Most recently, Rabaan has led youth-centered AI learning initiatives that address inequities in computing education and support critical engagement with data and algorithmic systems, alongside community data labs that center residents’ public transportation experiences and facilitate direct engagement with policymakers.