The Knock at the Door

The knock comes unexpectedly on a Tuesday afternoon, while a family is going about their routine. A caseworker in a thin coat is at the door, seeking entry into their home. The mother, who has lived in the same flat for nine years, is taken aback by the unannounced visit. She inquires about the person's identity and the reason for their presence, as this is the first investigation she has ever experienced.

The caseworker hesitates before explaining that the city received an unspecified report, leading to a required follow-up. This moment, replicated thousands of times annually in American cities, marks the start of a potential confrontation between a family and the child welfare system. The flag triggering this process is not an alert from a neighbor, teacher, or healthcare professional, but rather an algorithmic prediction generated from years of administrative data.

This risk-scoring system, based on complex calculations, assigns a numerical value to each family seeking assistance. A screener at the child protective services office examines this score, and if it crosses a predetermined threshold, a caseworker is dispatched to the family's residence. Throughout this chain of events, an unseen human decision maker plays a crucial role, but the family remains oblivious to the intricacies of the algorithm that influenced the outcome.

In the spring of 2026, American child welfare is increasingly driven by predictive systems that operate largely beyond the reach of procedural rights typically afforded to other consequential decisions in modern life. The system's opacity is starkly evident when compared to other domains, such as credit scoring or parking enforcement, which provide greater transparency and legal recourse.

The Markup, a non-profit investigative journalism organization, delved into the workings of the Administration for Children's Services (ACS) in New York City, revealing that the agency had been employing an algorithmic risk-scoring tool since 2016. This tool, initially introduced with limited public discussion, generated a score for every family entering the system and played a pivotal role in determining which cases warranted heightened scrutiny.

However, the research conducted by The Markup, based on internal documents and interviews with agency staff, demonstrated that the system disproportionately flagged Black and low-income families at rates exceeding the actual prevalence of confirmed maltreatment within those populations. Factors like postcodes, prior contact with public assistance programs, and neighborhood service density were not explicitly disclosed as race or class proxies but functioned as such in practice.

The ACS defended the tool, asserting that it was advisory and that human decisions ultimately prevailed, while simultaneously refusing to disclose the model's full feature set, weights, or technical documentation. The Markup investigation holds significance not only because it exposed an ongoing issue but also because it occurred in New York City, the largest city in the United States and the hub of one of the most extensive child welfare caseloads in the country.

The findings confirm that predictive systems are now pervasive in child welfare decision-making, eclipsing the need for transparent due process.