Should Indian Courts Use Algorithmic Risk-Assessment Tools?

An algorithmic risk-assessment tool, in this context, is software that takes inputs about a defendant — prior record, charge type, sometimes demographic or socioeconomic proxies — and outputs a score meant to predict the likelihood of flight risk or reoffending. Several US jurisdictions have used such tools to inform bail decisions for over a decade. The pitch is consistency: a score doesn't have a bad morning, doesn't know the defendant's family, isn't swayed by how the courtroom feels that day.

Abstract

This piece examines the evidentiary and constitutional problems that would confront any attempt to introduce algorithmic risk-assessment tools into Indian bail decisions, focusing on three issues: the opacity of the underlying models, the risk of encoding existing patterns of disadvantage as if they were neutral predictors, and the due process implications of a court relying on a score it cannot fully explain to the person it affects.

What these tools actually predict

A risk score is not a prediction of guilt, and it is not really a prediction of dangerousness either — it is a statistical correlation between certain inputs and past outcomes in a training dataset. If that dataset reflects, for instance, over-policing of particular neighbourhoods or communities, a tool trained on it will reproduce that pattern, dressed up as an objective risk figure. The tool doesn't know it's doing this. It has no concept of the past unfairness embedded in its own training data.

The opacity problem meets due process

Indian bail jurisprudence, including the framework under the Bharatiya Nagarik Suraksha Sanhita, requires a court to weigh specific, articulable factors — flight risk, tampering risk, severity of offence — and to give reasons. A proprietary algorithmic score that even the vendor won't fully disclose the workings of sits awkwardly against that requirement. A defendant has a right to challenge the basis on which their liberty is being restricted; a black-box score makes that right largely theoretical.

A number is not a reason. A court that relies on a score it cannot explain has not given a reason — it has outsourced one.

Where comparative experience points

• Multiple US studies found risk tools flagged Black defendants as higher-risk at disproportionate rates relative to actual reoffending outcomes, despite race not being a direct input.
• Several jurisdictions that adopted these tools have since restricted or abandoned them after independent audits found accuracy claims did not hold up under scrutiny.
• Where tools remain in use, the more defensible deployments treat the score as one input a judge can override, with the override rate itself audited, rather than as a determinative output.

What a defensible version would require

If Indian courts were to pilot such tools at all, the comparative record suggests any responsible version would need full model transparency available to defence counsel, mandatory disaggregated accuracy audits published on a regular schedule, and an explicit rule that the score is advisory, with a judge required to state independent reasons rather than deferring to it. None of that is currently built into how these tools are typically marketed or sold, which is itself a reason for caution before any pilot moves forward.