Adoptions and Algorithms: Why Artificial Intelligence Cannot Measure a Family

The use of Artificial Intelligence in child welfare and adoption systems promises to revolutionize the sector through algorithmic matching, but investigations a

Artificial intelligence possesses an extraordinary ability to sort through vast databases, identify hidden resources within bureaucratic systems, and propose matches based on recurring patterns, but it comes up against an insurmountable limit: it lacks the ontological and ethical tools to decide whether a person is "suitable" to fill the role of parent. The psychological, emotional, and financial evaluation of a prospective parent or a foster family is not a simple problem of mathematical optimization. It is an inquiry that touches on fundamental rights, socioeconomic inequalities, conditions of poverty, disability, ethnic origin and, above all, the future and safety of a child, making it essential that such a decision remain human, multidisciplinary, deeply reasoned, and legally challengeable.

The Illusion of Matchmaking: The Family-Match Case

The gap between technological promise and the highly delicate practice of adoption emerges with disarming clarity in the Associated Press investigation into the Family-Match algorithm. This computational tool, explicitly inspired by the mechanics of online dating applications, had been designed with the intent of assigning a "relational compatibility" score between children in the foster care system and prospective parents. The investigation, however, revealed extremely limited practical results in the states of Virginia and Georgia, bringing to light cases of referrals to families that were not in the least interested in or prepared for that type of match, and raising profound doubts about the effectiveness claimed by the software's promoters.

The failure of this approach is not limited to poor statistical accuracy, but is rooted in the very nature of the data processed. The Family-Match algorithm devoured and processed over 250 sensitive variables to calculate its compatibility indices. Among these parameters were highly intimate and potentially stigmatizing information, such as histories of past sexual abuse, criminal records, and the LGBTQIA+ identity of the minors involved. The central problem, in this scenario, is not solely the quality or efficiency of the proposed match, but the mass collection and unsupervised use of extremely intimate data within a software architecture that is opaque and inaccessible to public scrutiny.

Digital Surveillance, Bias, and the Crime of Poverty

The introduction of artificial intelligence into the U.S. child welfare system has so far followed a precise and troubling trajectory. An academic review based on the analysis of 50 publications concerning computational systems in child welfare demonstrates that the vast majority of implementations involve predictive models oriented exclusively toward risk calculation. Vital functions such as true relational matching, the integration of social workers' human perspectives, and the qualitative analysis of case notes remain, by contrast, dramatically under-integrated or entirely ignored by the software.

This focus on "risk" generates an algorithmic short circuit that penalizes the most vulnerable segments of society. An in-depth study of the Allegheny Family Screening Tool, a well-known U.S. predictive model, reveals an unmistakable structural bias: poor families or those belonging to ethnic minorities interact out of necessity to a much greater extent with public services, thereby accumulating a disproportionate amount of administrative data on their record. The algorithm, reading this mass of data and using past investigations as predictors for the future, ends up assigning these families systematically higher risk scores, amplifying a discriminatory vicious circle. The focal point is that poverty and financial hardship are not equivalent to parental neglect, but within public datasets they invariably end up appearing as a gigantic risk factor.

To mitigate this drift, researchers emphasize the need to sharply distinguish between two purposes of use: employing machine learning to identify and offer preventive resources to families in difficulty is profoundly different from using it to select which families to investigate or from which homes to remove children. The second hypothesis, given its devastating consequences, requires an evidentiary standard incommensurably higher than the first. Georgetown Law's Center on Privacy & Technology perfectly synthesizes these critical issues, denouncing the scant evidence of effectiveness of current systems, an almost nonexistent accountability, the power asymmetries between public entities and private software vendors, and the extreme danger in using socioeconomic proxies to assess parenting. Faced with this institutional picture, it becomes imperative to demand independent audits, transparent documentation, and concrete and accessible remedies for families unjustly penalized.

The Legal Value of Empathy: Rights and International Treaties

Delegating to a machine the formation or dissolution of a family clashes head-on with the architecture of international human rights. The United Nations Guidelines for the Alternative Care of Children (2009) establish a clear perimeter: every decision must be made strictly case by case and be based exclusively on the best interests of the child. This process requires solid legal safeguards, rigorous and qualitative assessment, and the direct involvement of multidisciplinary teams composed of highly qualified professionals. Such international criteria make it in fact impossible and illegitimate to justify a cold algorithmic score as the basis for an autonomous or semi-autonomous decision.

The UN Convention on the Rights of the Child further reinforces this paradigm. In the context of adoption, the best interests of the minor cannot be downgraded to just one of many variables to be calculated: it is the primary and determining criterion, which must include, wherever possible, the active participation and listening to the voice of the child themselves. The minor in search of a family cannot and must not be reduced to a mere "output" of matching software, but must be treated as a person who is a bearer of rights, of an unrepeatable story, and of a will of their own.

The Council of Europe recently intervened to legally lock down the interaction between AI and fundamental rights with the Framework Convention on Artificial Intelligence of 2024, the first binding international treaty on the matter. The document requires that the use of AI respect human dignity and autonomy, guarantee equality and non-discrimination, and imposes human rights impact assessments and strict preventive measures throughout the entire lifecycle of the system. The regulatory message is unequivocal: a system tasked with assessing parenting capabilities cannot in any way operate as an inexplicable black box.

The Final Word

Contemporary debate must take a leap in maturity. The central question that policymakers and social workers must ask is not the science-fictional and illusory "can artificial intelligence predict whether we will be good parents?", but rather a series of pragmatic and systemic questions: what specific signals is the algorithm using to decide, who has labeled these signals as risk indicators, and through what channels can a citizen challenge a verdict generated by the machine? A low income, a past in social services records, a disability condition, or a complex family history are not and will never be tools capable of measuring love, aptitude for care, or the real ability to raise and educate a child; they can, however, turn into inexorable risk proxies if the artificial intelligence model is trained on historical data already deeply marked by systemic inequalities and injustices.

An algorithm can excel at sifting through bureaucratic files, can suggest a territorial resource or hypothesize a potential meeting, but it does not have the moral, legal, or cognitive authority to decide the intrinsic value of a family. When the life, future, and emotional stability of a vulnerable child are at stake, the final word must always and unavoidably belong to flesh-and-blood people, responsible professionals, capable of explaining their reasoning transparently, able to listen to unexpressed needs, and bound to answer personally, both ethically and legally, for their choices.

Bibliographic References and Sources

• Does an AI tool help boost adoptions? Key takeaways from an AP Investigation — Associated Press, 2023
• Inspired by online dating, AI tool for adoption matchmaking falls short for vulnerable foster kids — Associated Press, 2023
• A Human-Centered Review of Algorithms used within the U.S. Child Welfare System — CHI 2020
• A case study of algorithm-assisted decision making in child welfare — PMLR, 2018
• Using machine learning to predict future foster care admission — PMC, 2025
• Data and Automation in Child Welfare — Georgetown Law, Privacy & Technology Center
• Guidelines for the Alternative Care of Children — United Nations, 2009
• Convention on the Rights of the Child — United Nations
• The Framework Convention on Artificial Intelligence — Council of Europe, 2024

Article by the Editorial Staff of La Bussola dell'IA