Los Angeles County Project AURA

Project AURA was a predictive risk model that the analytics firm SAS built for Los Angeles County's Department of Children and Family Services (DCFS). It scored each child in a referral, a report of suspected abuse or neglect, from 1 to 1,000, using records from several agencies. How SAS's model produced each score was proprietary.

What the test found

AURA was trained and tested on past cases, never on a live one. The test compared its flags with child deaths, near-fatalities, and critical incidents from 2011 and 2012. The sources do not define a critical incident.

The sources report results at one high-risk cut: the score above which AURA flagged a child. At that cut, AURA correctly flagged 171 children who had one of those worst outcomes. It also flagged 3,829 children who did not. Those are false flags, or false positives. That is about 22 false flags for every correct one.

DCFS's own public-affairs director confirmed on the record a false-positive rate of no less than 95.6 percent. That is the share of flagged children who did not have one of those outcomes.

What else the sources say about the model

The algorithm was proprietary to SAS, and no one outside SAS knows exactly how it works. WitnessLA says SAS ran the check of what actually happened to the families.

How it ended

AURA was never used on a single live case. In May 2017 the county's Office of Child Protection reported to the Board of Supervisors that DCFS was no longer pursuing it. By then the false-positive results had been known for nearly two years.

A 2017 commentary by Richard Wexler in WitnessLA says that news was buried on page 10 of the report. It says it is not clear what finally prompted DCFS to drop AURA. Perhaps, it says, it was the report's point that all those false positives would further overload the system.

More likely, the commentary says, it was a State of California initiative to come up with a "better" predictive analytics model.

The same report called for strict standards before considering the use of predictive-analytics models. They include understanding how racism and other biases may be embedded in systemic data.

The background

DCFS contracted SAS around 2013. AURA's development unfolded amid intense scrutiny of DCFS after the May 2013 death of eight-year-old Gabriel Fernandez in Palmdale.

What came after

In 2021 the county launched a separate, in-house tool, the Risk Stratification Pilot. Unlike AURA, it is not proprietary, and it draws only on the county's own records. It deliberately excludes race, ethnicity, and geography.

Why this case matters

The case file reads AURA's test as exposing a base-rate problem that no accuracy figure could hide. A base rate is how common an outcome is. When the outcome is rare, even a fairly accurate screener flags many children who will not have it.

The county was evaluating a model it could not open. What it had was an outcome count, not a look at how the model worked. The decision also came slowly.

The case file reads refusing deployment as a governance decision in its own right, not a failure to ship. The question this case asks is not how to run the tool safely. It is which controls decide, before go-live, whether it runs at all.

Where the facts come from

The sources include 2015 reporting in The Imprint, 2015 reporting by KPCC/LAist, and a 2015 Child Protective Services Defense article. They also include Richard Wexler's 2017 commentary on the NCCPR blog and in WitnessLA. The Wexler pieces and the Child Protective Services Defense article are advocacy.

Later sources cover the Risk Stratification Pilot, from the county, DCFS, and the Children's Data Network at the University of Southern California. An NBC Los Angeles timeline covers the Gabriel Fernandez case.

What this network is drawn from

This network follows the pattern the sources describe. It is not a reconstruction of the actual tool. AURA never ran on a live case, so the network shows the shape it would have had in use.

The retrospective test used de-identified cross-agency records, with the details that identify a person removed.

What the available tools can and cannot address

A failure pathway is a link between two parts of the network, where a mistake made by one part can be passed on to the other. A closed pathway is one that mistakes stop passing along. The work along it may go on.

This case has a budget of 10 units. Before any tool is used, six failure pathways are open. The network is at a tipping point, where mistakes are on the edge of building on one another.

Explore (No Targets) sets no targets. There, and under Service Targets Only, a single tool is enough to keep mistakes from building on one another. Three do it for 2 units each: Assign a challenger, Escalate checks, and Mark AI-written records. Understand the system also does it alone, for 3 units. So does Upgrade model at its stronger setting, for 5 units, with lingering effects on.

Under Service Targets Only, each of those also meets the service target. That target asks for the risk score to be helping the investigators' work. Hundreds of other combinations within the budget do too.

Under Service and Safety Targets and All Governance Targets, the targets can be met in exactly one way within the budget. Both levels also ask you to close every failure pathway. The one way uses four tools at their standard settings, for all 10 units. They are Mark AI-written records, Escalate checks, Check with a second model, and Store less data.

Each of the four closes pathways the others leave open. Mark AI-written records labels the scores logged in the records. Mistakes then stop passing from the records and the cross-agency data into the score, and from the records to investigators.

Escalate checks has investigators check the score more closely when monitoring flags trouble. Mistakes then stop passing along Score to investigators. Store less data stops them passing along Investigation findings to the records. Check with a second model stops them passing along One model for every referral.

With the budget set aside, every combination that meets these two levels still includes all four. At All Governance Targets, Store less data works at reduced strength unless Understand the system is also on. It still does enough here.

The tools closest to the county's decision not to deploy AURA close no pathway here. Require sign-off, Gate vendor updates, and Review on schedule change none. Assign a challenger keeps mistakes from building but closes none. Pause AI on alarms stops investigators getting the score, but the work then falls behind.

This network shows the score already in use. So no tool here takes it out of the investigations the way the county's decision did.

More is not better here. Every tool at its strongest setting at once costs 41 units, about four times the budget. It closes every failure pathway and keeps mistakes from building. But the risk score then no longer helps the investigators' work enough, so this misses the targets at all three levels that set them.