Rewriting the Algorithm: How a Pattern-Based Approach Offers a Human-Centered Alternative to Risk Prediction in Child Welfare
By David Mandel, CEO and Founder, and Ruth Reymundo Mandel, Chief Business Development Officer and Credible Expert, Safe & Together Institute
In a 2024 article published by the American Bar Association, legal scholars and child welfare advocates examined the rising use of algorithmic decision-making tools in child protection. Titled “Algorithmic Decision-Making in Child Welfare Cases and Its Legal and Ethical Challenges,” the piece outlines the potential dangers of relying on predictive tools to determine child welfare outcomes. Although these tools are often framed as neutral and efficient, they promise far more than they deliver—especially for families impacted by domestic abuse.
As jurisdictions across the globe explore integrating artificial intelligence into child protection systems, we must ask: What kind of decision-making do families deserve? How do we ensure it is fair, informed, and ethical?
We argue that the solution lies not in more opaque algorithms but in grounded, human-centered frameworks built around actual behaviors, lived realities, and systems accountability. The Safe & Together approach, developed through decades of fieldwork, survivor insight, and practitioner experience, offers this alternative by centering on perpetrator patterns, documenting multiple pathways to harm, and scaffolding professional judgment—not replacing it.
Flawed Foundations: The Legacy Risk Models Informing Predictive Tools
Before examining predictive analytics, we must scrutinize the risk assessment models they are built upon. Most traditional child protection frameworks were never designed with domestic abuse at their core. As a result, they miss key dynamics like the impact of the perpetrator on the parenting of the adult survivor, the use of children as tools of control, interference with healthcare or schooling, and legal systems abuse, amongst other issues.
Because most child protection practice models and related risk frameworks and data systems do a poor job identifying behavioral patterns of coercive control or isolating perpetrator behavior from survivor actions, the risk and danger to children created by domestic abuse remains misclassified or completely invisible. This not only endangers survivors and children—it warps the very data we use to make decisions. And it is at odds with the lived experience of practitioners—many of whom say that domestic abuse is present in 80% of the child protection caseloads—and the lived experience and critical needs of families (a recent study showed that almost 40% of Australians grow up in households with domestic abuse).
Existing risk frameworks focus heavily on static data points, compliance measures, and incident-based thinking. They often do not include critical concerns like ongoing coercive control or the cumulative impact of perpetrator behaviors on the holistic functioning of children and families. Worse, they almost universally exclude adult survivor safety from child risk assessments—a glaring oversight, given that the harm and fear experienced by the survivor often reflect the child’s daily reality. That gap becomes even more stunning when you consider how central survivors’ protective efforts are to children’s safety and well-being.
The absence of adult survivor safety and self-determination from the child risk equation leads to flawed decisions. It contributes to punitive responses like child removals and misclassifies protective efforts as noncompliance. And these distortions are compounded by gendered assumptions that scrutinize mothers while rendering fathers who use violence invisible.
Predictive Risk Models Compound Old Mistakes
Rather than correcting these flaws, predictive analytics can magnify them. Risk scores are generated from historical data sources—arrest records, welfare involvement, housing instability—that may reflect systemic racism, poverty bias, and documentation that often fails to distinguish between the behaviors of perpetrators and survivors.
Families who are marginalized are disproportionately flagged—not because they are at higher risk but because of decades of biased data entry. Survivors of domestic abuse may be misclassified as risks because their attempts to seek help have left a digital trail. Children’s trauma responses can be misunderstood. And most critically, perpetrators’ patterns of behavior often remain untracked, undocumented, and therefore invisible.
A Pattern-Based Alternative: Making the Invisible Visible
A pattern-based approach avoids many of these issues by focusing on actual behaviors related to a specific family and situation. The analysis of risk focuses follows the simple concept that the better we understand the past behavior of a person who uses violence and control, the better we can become at anticipating how they might act in the future. And on top of this, as opposed to risk factor assessments that look for “tick box” factors, a behavioral focus helps describe harm and gives better contextual understanding of the situation.
When we use this approach, we ask different questions:
What has the person choosing violence done as a parent?
How have those behaviors harmed the children, partner, and family functioning?
How has the other parent protected the children in the face of those behaviors?
How have systems responded—or failed to respond—to these dynamics?
These are not speculative questions. They can be answered through observation, interviews, and documentation. They offer a framework that is transparent and, most importantly, anchored in real behaviors.
By mapping patterns of perpetration, child protection agencies can shift from assigning risk based on poverty or past involvement to identifying risk based on present-day danger and parenting choices. This changes everything: from the focus of the case plan to who is held accountable to how we support survivors and their children.
A Better Data Culture
This approach not only improves practice, but it also improves data collection and analysis. We are not anti-data. But we must ask: What data points are we collecting? What meaning do we assign to them? And what do we do with that data?
A behaviorally grounded model tracks:
Parental perpetrator behavior patterns across time and relationships
Survivor protective efforts in the context of the perpetrator’s pattern and the systems’ responses
Child functioning and perspectives in context of both parents’ behaviors
Systems responses efforts directed at the perpetrator, focused on the parent causing danger and harm
This is the data that can drive real practice improvement. It tells us who did what to whom, how it affected family life, and what role the system played. And it allows for public transparency, internal reflection, and collaborative solutions.
The Future We Choose
As CW360°’s 2025 issue on AI and child welfare makes clear, predictive analytics are not neutral, and historical data is victim-skewed and gender-biased. Their use raises fundamental questions about fairness, bias, accuracy, and effectiveness. Child welfare systems around the world must choose: Will they continue relying on opaque, siloed tools built on fragmented historical data that often ignore the parent causing harm—or will they reinvest in behaviorally grounded practices that reflect the real dynamics families are navigating?
A behavior-based approach—rooted in professional expertise, relationship-based work, and domestic abuse–informed thinking—offers a path toward consistent, ethical, accurate, and accountable decision-making. It aligns with our values. It honors survivors. It protects children. And it strengthens our workforce.
It’s time to move beyond the algorithm.
