NSAG · Module M2 · Legal & Justice
AI Legal Navigation
Standalone deployment retired
Algorithmic risk assessment tools now influence bail, sentencing, and parole for approximately two-thirds of the US population. Most of the people affected don't know these tools are used. Most institutions using them have no independent evaluation of their performance.
What this address is
This hostname served a standalone copy of NSAG module M2. That copy was retired on 15 August 2026, and the page you are reading replaced it. The deployment stays online so that links already published against it keep resolving, and so that anyone arriving here is sent to the material that is still maintained.
The module's current scope, its evidence base, and its release status are published on the NSAG hub at nsag-site.vercel.app/m2. Where this page and the hub disagree, the hub is correct.
What the module examines
Risk assessment algorithms are sold as objective replacements for subjective judicial bias. The evidence does not support this framing. Dressel and Farid (2018) demonstrated that COMPAS performs no better than untrained lay prediction while producing racial disparities in false-positive rates. Berk et al. (2024) showed that the bias originates in the training data — arrest records encoding decades of racially disparate policing — making technical corrections insufficient. The governance response is to build the institutional infrastructure that makes these tools accountable.
M2 sits in the Legal & Justice group of the framework.
What the assessment measured
The module organised a structured self-assessment across six governance dimensions:
- 1AI Tool Transparency & Explainability
- 2Bias Audit & Demographic Impact Assessment
- 3Human Oversight & Override Protocols
- 4Plain-Language Communication to Affected Parties
- 5Accountability & Error Correction
- 6Community Input & Governance Structures
Each dimension was described against tiers running from early stage up to the fully implemented tier the framework calls PIONEERING, with observable criteria written for each level, so that an institution could locate its own arrangements rather than receive a score. It was a self-assessment framework for institutional reflection, and never a validated instrument, an audit, an accreditation, or a compliance determination.
Who it was written for
Court innovation offices · Legal aid tech teams · Public defender AI tool governance · Access-to-justice organizations · Bar association AI policy committees
And anyone a risk score is run on. Most people scored by these tools never know. If a decision about your bail, sentence, or parole runs through an algorithm, this governance is your recourse.
Why the standalone deployment was retired
The fifteen modules were first published as fifteen separate deployments. Scope, evidence, and release status then had to be maintained in fifteen places, and they drifted apart. The hub now holds one canonical page per module, and these fifteen addresses point at it.
Assessment collection is paused across all fifteen modules. The published operations matrix records the same position for every one of them: the canonical route is reachable, collection is paused, and advisory work is delivered by a person rather than by automated scoring. This page is a static record. It carries no forms and collects nothing.