When child-protection supervisors could see a risk score, repeat reports fell 15% — and foster placements nearly doubled
A Pennsylvania trial let human supervisors see an algorithm's five-level summary of information already in the case file. The score did not make decisions, but access to it changed which referrals received the most intensive response.
Child-protection supervisors received reports about possible abuse or neglect. For some family referrals, a computer-generated five-level complexity rating was visible alongside the usual case file; for others, the rating was hidden. Supervisors and investigators retained discretion over agency decisions, including whether to ask a court to place a child in foster care.
The score changed the cases labeled most complex
Workers did not simply intervene more in every case. The biggest changes appeared among referrals the model labeled Very High complexity. In that group, the estimates imply that ongoing services rose from about 9% to 19%, while substantiated referrals rose from about 34% to 49%.
The 15% reduction in repeat reports applies to the full study sample. The apparent decline within the Very High group was too uncertain to distinguish from chance.
The score appears to have redirected attention, not replaced human judgment. The experiment cannot tell whether the summary provided new insight, increased workers' confidence or simply made intervention feel institutionally supported.
What this study cannot tell us
A repeat report means another child-protection referral involving the child within three months. It is not direct proof that abuse or neglect occurred — or that children were safer when reports declined. More surveillance, less surveillance and contact with support services can all affect reporting.
The trial covered one Pennsylvania county and 34 investigators over 14 months. Researchers did not detect wider racial disparities when the score was visible, but those subgroup estimates were too imprecise to prove equal effects. This is a July 2026 NBER working paper and has not been peer reviewed.