The Arnold Venture's Predictive Risk Assessments Found to Have no Predictive Abilities at All
For how long have we been hearing that risk assessments are the answer to a more fair system? How long have we heard to trust the experts and the algorithms? Well guess what?
For how long have we been hearing that risk assessments are the answer to a more fair system? How long have we heard to trust the experts and the algorithms? Well guess what? A new Harvard Law School Access to Justice Lab study, commissioned about a decade ago by Arnold Ventures to evaluate the effectiveness of the PSA — a widely used pretrial risk assessment tool — in Dane County, Wisconsin, as well as sites in Iowa and Utah, has finally been completed. And guess what...the risk assessment didn't make the system more fair. It didn't make our communities more safe. What did it do? Absolutely nothing.
The sales pitch that Arnold Ventures uses for their PSA is that it uses factors such as age and criminal history to predict the likelihood that a defendant will fail to appear in court or commit new (including violent) crimes. In other words, it predicts the future. The goal of providing the tool to decision-makers was to improve public safety, reduce failures to appear, lower new criminal activity, reduce pretrial jail time, and avoid worsening racial disparities. A sales pitch that not only seems unbelievable, but also never seems to materialize.
According to a recent article in the Cap Times, here are the key findings from the Harvard study:
- The PSA produced no statistically significant improvements on any of the five main hoped-for outcomes.
- It also did not make racial disparities or incarceration rates worse.
- Pretrial jail stays were essentially the same whether the PSA was used or not.
- Across the study sites, the tool was neutral: “not making anything worse but not making anything better.”
Lead researcher Jim Greiner concluded that because traditional bail itself only affects outcomes for a very small fraction of defendants, giving decision-makers advice about when to impose bail is largely ineffective. He recommended that Dane County consider redirecting the money spent on producing the PSA to other approaches (or simply saving it) and suggested future research focus on monitoring and services rather than risk assessment tools.
This study fits a broader pattern of pretrial risk assessment tools not only failing to deliver the transformative results their proponents promised but failing to deliver anything worthwhile at all.
For years, advocates of “data-driven” pretrial reform (including Arnold Ventures, which funded the development and early rollout of the PSA) argued that algorithmic risk assessments would allow jurisdictions to safely reduce pretrial detention, improve court appearance rates, and enhance public safety while reducing racial disparities. The tools were presented as scientific, objective alternatives to traditional bail decisions.
In practice, rigorous evaluations, including this randomized study, have repeatedly shown limited or no measurable benefits on the key outcomes that justified their adoption. The tools often function more as expensive administrative processes than as effective decision-support systems. Because the underlying assumption (that better prediction of risk will meaningfully change judicial behavior and defendant outcomes) has not held up under careful testing, jurisdictions continue spending resources on instruments that neither significantly improve safety nor reduce unnecessary jail use.
The Dane County results reinforce a growing body of evidence that pretrial risk assessments have largely failed to live up to their marketing. Instead of producing clearer, safer, or fairer release decisions at scale, they have often become another layer of bureaucracy whose costs are easier to measure than their benefits. The researcher’s own conclusion, that the money spent on the PSA might be better used elsewhere, underscores the gap between the original reform vision and the actual results after more than a decade of implementation and study.
