Fair Go Justice — Legislative-Grade AI Reference

AI TECHNOLOGY IN JUSTICE

A comprehensive, plain-language guide to every major category of artificial intelligence deployed in — or proposed for — Australia's legal system. Written for Senators, lawyers, journalists, and the public who deserve to understand the technology making decisions about their lives.

YOU CANNOT GOVERN WHAT YOU DON'T UNDERSTAND

In April 2026, the Federal Court of Australia issued a formal warning after lawyers submitted AI-generated fake case citations in real proceedings. Algorithms are determining welfare eligibility, bail decisions, aged care allocations, and family court outcomes — with no mandatory disclosure, no right to explanation, and no independent oversight.

This page exists so that every Australian — from a Senator on the Legal and Constitutional Affairs Committee to a self-represented litigant preparing for court — has access to accurate, legislative-grade information about the technology shaping their rights.

🤖
Category 01
LARGE LANGUAGE MODELS (LLMs)
⚠ High Risk — Court Citation Integrity

What They Are

Large Language Models (LLMs) — including ChatGPT, Claude, Gemini, and Llama — are AI systems trained on billions of text documents that generate human-like text in response to prompts. They predict the most statistically likely next word based on patterns in their training data. They do not reason. They do not look up information. They generate plausible-sounding text.

The Justice System Risk

LLMs hallucinate — they generate false information with complete confidence. In legal contexts, this means fabricated case citations, invented statutory provisions, and non-existent judicial authorities submitted to courts as real. In Australia, 73+ documented cases occurred before the Federal Court issued its warning.

Current Justice Uses
Legal research assistance, document drafting, court submission preparation, victim communications (Services Australia), tribunal decision summaries
FGJ Required Safeguards
Mandatory disclosure of LLM use in any court document; citation verification requirement; human lawyer sign-off on all AI-assisted submissions

Key Australian case: Luck v Secretary, Services Australia [2025] FCAFC 26 — first Federal Court case explicitly addressing AI-assisted decision-making (February 2025).

📝
Category 02
NATURAL LANGUAGE PROCESSING (NLP)
⚠ Medium Risk — Automated Decision Classification

What It Is

Natural Language Processing (NLP) is a branch of AI that enables computers to read, analyse, and interpret human language — including legal documents, transcripts, affidavits, and correspondence. NLP classifies text, extracts key information, and identifies patterns at scale.

The Justice System Risk

NLP is used in automated triage systems to classify legal claims, identify risk categories for bail decisions, and process welfare applications. When NLP misclassifies — due to training data bias or domain mismatch — it produces systemic discrimination at industrial scale without any single human making a discriminatory decision.

Current Justice Uses
Bail risk assessment tools, child protection case classification, welfare eligibility triage, court document management, transcript analysis
FGJ Required Safeguards
Mandatory bias audit of all NLP classification systems used in justice contexts; human review of any NLP-assisted adverse determination; disclosure to affected parties
⚙️
Category 03
AI AGENTS & AUTOMATED DECISION SYSTEMS
⚠ High Risk — Autonomous Rights Decisions

What They Are

AI Agents are systems that take autonomous actions in the world — not just generating text, but making decisions, triggering workflows, and acting without human input at each step. Automated Decision Systems (ADS) apply rules and AI outputs to make administrative decisions at scale.

The Robodebt Lesson

Robodebt was Australia's first mass-scale ADS disaster. An automated income-averaging algorithm sent 433,000 debt notices without legal authority. $548.5 million was repaid. Lives were destroyed. The Royal Commission found the system was designed to avoid human review — because human review would have identified the illegality immediately.

In November 2025, an aged care allocation algorithm denied services to 18,000+ eligible recipients with no human review pathway and no appeal mechanism. Robodebt did not end. It evolved.

Current Justice Uses
Welfare debt assessment (former Robodebt, current systems), aged care allocation, visa processing, parole condition monitoring, child protection risk scoring
FGJ Required Safeguards
Human-in-the-loop mandatory for all rights-affecting decisions; right to explanation; independent algorithmic audit; retroactive review since January 2020
👁️
Category 04
COMPUTER VISION & BIOMETRIC SYSTEMS
⚠ High Risk — Evidence and Identity

What It Is

Computer vision enables AI to interpret images and video — including CCTV footage, facial recognition, body camera analysis, and document authentication. Biometric systems use facial features, gait, voice, and other biological characteristics to identify individuals.

The Justice System Risk

Computer vision is now used in evidence analysis — identifying individuals in CCTV footage, assessing vehicle registration in traffic offences, and verifying document authenticity. Facial recognition systems have documented error rates of 10–35% for people of colour. In criminal justice contexts, these errors are not statistical abstractions. They are wrongful convictions.

Current Justice Uses
CCTV evidence analysis, facial recognition for suspect identification, document fraud detection, deepfake detection in evidence (emerging), border security biometrics
FGJ Required Safeguards
Mandatory disclosure of computer vision use in criminal proceedings; independent validation of any biometric evidence; prohibition on sole reliance on facial recognition for identification
🔗
Category 05
BLOCKCHAIN & DISTRIBUTED LEDGER TECHNOLOGY
🔵 Emerging — Evidence Integrity Solution

What It Is

Blockchain is a distributed, tamper-resistant ledger — a record of transactions or events that cannot be altered after the fact without detection. Unlike centralised databases, blockchain records are verified by multiple independent nodes, making tampering computationally infeasible.

The Justice Opportunity

Blockchain is the technological solution to the evidence tampering problem at the heart of Pillar 3. A blockchain-based chain-of-custody system would create an unbreakable digital record from the moment evidence is collected — time-stamped, independently verified, and impossible to alter without detection. This is not theoretical. Estonia has used blockchain for national legal records since 2012.

Proposed Justice Uses
Digital chain-of-custody for all physical and digital evidence; immutable court record archive; transparent petition and submission verification; judicial appointment records
FGJ Reform Alignment
Pillar 3 — Evidence Integrity: blockchain as mandatory chain-of-custody standard; Pillar 4 — AI Accountability: algorithmic audit trails on distributed ledger
🇦🇺 Australian AI in Justice — Current Landscape

WHO IS ALREADY WORKING ON THIS

Australia is not starting from zero. A growing ecosystem of organisations is building responsible AI tools for justice contexts — and they should be partners in, not obstacles to, reform:

WHERE THE WORLD IS GOING

FrameworkJurisdictionKey ProvisionStatus
EU AI ActEuropean UnionProhibits AI in criminal justice without human oversight; mandatory transparency for high-risk usesIn force 2024
UNESCO AI Ethics Recommendation193 nationsHuman rights-centred AI; right to explanation; algorithmic accountabilityAdopted 2021
OECD AI PrinciplesOECD members incl. AustraliaTransparency, explainability, human-centred values — Australia is a signatoryActive
UK Justice AI FrameworkUnited KingdomCourt rules for AI disclosure; Law Society AI Ethics guidelinesPublished 2023
NCSC AI Readiness FrameworkUnited StatesGovernment AI readiness standards — model for Australian adoptionPublished 2024
AI Accountability in Justice Act 2026Australia — ProposedBinding AI disclosure + human-in-the-loop + right to explanation + oversight panelFGJ Proposal — 2026

UNDERSTAND THE TECHNOLOGY. DEMAND THE SAFEGUARDS.

Every Australian has the right to know when AI is making decisions about their life — and the right to challenge it. Join the movement demanding that right become law.

Read Pillar 4 → Join the Movement →
Field Evidence — NLP & AI in Live Legal Aid Deployments

Real World Results: NLP Transforms Legal Aid Delivery

The following results are drawn from the OECD's 2025 report on AI in justice administration — the gold standard reference for policymakers and Senate committee submissions globally.

🇧🇷 Brazil — Appeal Evaluation (NLP)
44 min
Human-only
2–3 sec
With AI
98.5% time reduction
🇵🇪 Peru — Protection Measures (NLP)
3 hrs
Human-only
40 sec
With AI
99.6% time reduction
🌐 Global — Lawyer Time Savings (GenAI)
32.5
Working days saved per lawyer per year
6–20% weekly time savings (Everlaw, 2024)
🇦🇺 Australia — Justice Connect NLP
245
Lawyers made 90,000+ annotations
across 9,000+ NLP training samples
Lawyers trained the AI — not replaced by it

Source: OECD — Governing with Artificial Intelligence (2025): AI in Justice Administration and Access to Justice

This report forms the primary international benchmarking reference for the AI Accountability in Justice Act 2026. It is the same framework cited by the European Commission and the UN Office on Drugs and Crime in their respective AI justice guidance documents.

OECD AI Principles — Compliance Mapping

The AI Accountability in Justice Act 2026 maps directly to all five OECD AI Principles and Australian Government AI Policy v2.0. This is not speculative policy — it is alignment with international standards already in force.

OECD AI Principle FGJ Act 2026 Measure Status
1. Inclusive Growth & Well-Being Universal legal access mandate; 22,000 new positions; regional expansion coordinators COMPLIANT
2. Human-Centred Values & Fairness 3-tier human oversight; Ethics Review Committee; ATSI cultural protocols; bias audit requirements COMPLIANT
3. Transparency & Explainability Mandatory public reporting; algorithm audit logs; explainable AI requirement; annual independent auditor report COMPLIANT
4. Robustness, Security & Safety 5-step incident response; sovereign infrastructure requirement; Privacy Act 1988 + 13 APPs enforcement; PIA mandatory COMPLIANT
5. Accountability Named accountability structure; 5 specific roles; Senior AI Review Counsel sign-off; Parliamentary reporting obligations COMPLIANT

3-Tier Human-in-the-Loop Oversight Architecture

TIER 1

Client Level

AI output presented to client with plain-language explanation. Client consent required before any AI-assisted advice is relied upon. Voluntary participation framework — no client is forced into an AI-assisted process.

TIER 2

Lawyer Level

Qualified lawyer reviews, modifies and certifies every AI output before it reaches a client or court. The AI is a research and drafting tool — the lawyer makes every consequential decision. Professional obligations fully intact.

TIER 3

Institutional Level

Senior AI Review Counsel holds final sign-off authority for high-stakes determinations. National AI Justice Board oversight. Algorithm audit logs retained. Annual public reporting to Parliament.

Mandatory Training — Lawyers

16 Hours

CPD-accredited curriculum covering: AI fundamentals for lawyers (4h), Ethical oversight and bias detection (4h), Practical AI tools and workflow (4h), Privacy, confidentiality and liability (4h). Required before deployment access granted.

Mandatory Training — Support Staff

8 Hours

CPD-accredited curriculum covering: Understanding AI tool outputs and limitations (2h), Data handling, privacy and consent (2h), Escalation pathways and when to flag for lawyer review (2h), Incident identification and reporting (2h).

5-Step Incident Response Protocol

1
Identify & Flag — Any lawyer, support staff or client may flag a suspected AI error or bias incident. Mandatory escalation within 2 hours.
2
Immediate Isolation — Affected AI process suspended within 4 hours pending review. No further outputs from the affected module until cleared.
3
Senior Review Counsel Assessment — Senior AI Review Counsel assesses all affected determinations within 24 hours. Human decision substituted where required.
4
Technical Audit — Technical Standards Panel root-cause investigation within 72 hours. Corrective action plan published.
5
Public Disclosure — Material incidents disclosed in next annual Parliamentary report. Major incidents disclosed publicly within 30 days. No suppression. No exceptions.

Privacy & Data Sovereignty Framework

Privacy Act 1988
Full compliance + all 13 Australian Privacy Principles enforced at system level
Sovereign Infrastructure
No client data stored offshore. Australian cloud/data centres only.
Mandatory PIA
Privacy Impact Assessment required before any AI system deployment
AU Gov AI Policy v2.0
Full alignment with Australian Government AI Ethics Principles — all 8 mapped