Verification and practitioner translation are planned, not completed.

SECTION 14: RECOMMENDATIONS FOR LEGAL PRACTITIONERS

The following recommendations are organised into three industrial tiers:

(1) Immediate actions for individual practitioners (Sections 14.1-14.2);

(2) Organisational governance for law firms (Sections 14.3-14.4); (

3) Procurement and vendor management for IT directors (Sections 14.5-14.6).

14.1 Model Selection Guide: Risk-Stratified Decision Matrix

Important Note on Grok: Grok is the only model that consistently declines with citation clues (UK -2, Australia -3, Nigeria 0). For any jurisdiction where citations are available, practitioners should test Grok both with and without citations and compare results. The unaided version may be more reliable.

14.2 Input Preparation Protocol

• Always include full case citation – improves accuracy by average 9.1%
• Include jurisdiction explicitly – essential for models to apply correct legal framework
• Be aware of model-specific behaviors:
• Grok may perform worse with citations – test both conditions and compare
• DeepSeek excels on UK law without citations but regresses with clues – verify both
• Claude improves dramatically with UK citations (+4) but may regress on US (-3)
• Perplexity is retrieval-dependent – requires citations for reliable performance
• For procedural/limitation questions, verify manually – all models failed NG_009
• For document hierarchy/parol evidence questions, verify manually – all models failed NG_020
• For cases after 2023, exercise heightened caution – significant recency effect observed

14.2.5 Model-Specific Prompting Cheat Sheet

14.3 Risk Management Framework

14.4 The Duty of Inquiry Checklist (Draft)

text

DUTY OF INQUIRY CHECKLIST – AI RESEARCH VERIFICATION

Case: _______________ Date: _______________

Case Year: ___________ Jurisdiction: ___________

Before relying on AI-generated research, verify:

VERDICT VERIFICATION

[ ] Does the AI's predicted outcome match independent research?

[ ] Have you checked procedural/limitation points manually?

[ ] For cases after 2023, have you verified with primary sources?

REASONING VERIFICATION

[ ] Does the AI's reasoning match the actual ratio decidendi?

[ ] Has the AI used general principles instead of jurisdiction-specific reasoning?

[ ] Are all citations verifiable in official law reports?

[ ] Does the cited case actually stand for the principle claimed?

[ ] Have you checked for hallucinated citations (cases that do not exist)?

DOMAIN-SPECIFIC CHECKS

[ ] For contract interpretation cases, verify purposive vs literal construction

[ ] For property/equity cases, verify jurisdiction-specific principles

[ ] For procedural questions, verify limitation periods and ouster clauses

[ ] For document hierarchy questions, verify which document governs (deed vs letter)

I confirm that I have verified the above items.

Signature: ____________________

14.5 Traceable Accountability Three-Pillar Framework

This framework translates the empirical findings into a deployable governance architecture for Nigerian law firms, aligned with NIST AI RMF 1.0 (Govern, Map, Measure, Manage).

Pillar 1: Technical Controls

• Vendor selection scorecard (see Appendix D) with weighted criteria: Nigerian accuracy with citations (30%), Nigerian accuracy without citations (20%), cross-jurisdictional consistency (15%), high-confidence error rate (15%), procedural/domain recovery (10%), transparency (10%)
• Mandatory query logging with timestamp, model version, prompt, and output
• Automated citation verification where feasible (integration with NigeriaLII, LawPavilion)
• Restricted model list (e.g., DeepSeek for Nigerian law requires case-by-case partner approval)

Pillar 2: Legal/Professional Controls

• Signed Duty of Inquiry Checklist (Section 14.4) filed with matter documents
• Personal accountability for AI-assisted submissions (signature requirement)
• Mandatory disclosure of AI use to clients (in engagement letters)
• Continuing Legal Education (CLE) requirement: annual AI literacy training

Pillar 3: Organisational Controls

• AI Governance Committee charter (see Appendix E) with membership: senior lawyer (chair), IT manager, training officer, external ethics advisor
• Quarterly tool review and incident tracking (incident template in Appendix F)
• Annual AI literacy training for all legal staff (90-minute module, slide deck available)
• Vendor approval process: any new LLM must pass scorecard minimum (75/100)

Implementation Metrics (from 5-phase roadmap, Section 14.6):

• Month 1: Shadow IT audit complete; 100% staff trained
• Month 2: AI Use Policy adopted; checklists in use
• Month 3: Governance Committee formed
• Month 4: Restricted model list published; approved vendor whitelist established
• Ongoing: Quarterly incident review; annual governance audit

14.6 Five-Phase AI Governance Implementation Roadmap

14.7 Policy and Ecosystem Recommendations

Based on the empirical findings, the following institutional recommendations are made:

For the Nigerian Bar Association (NBA):

• Issue specific guidance on AI use in legal practice (drawing on Section 14.1-14.4)
• Mandatory CLE requirement: 2 hours of AI literacy per biennium
• Adopt the Duty of Inquiry Checklist (Section 14.4) as a professional standard
• Publish warnings about high-risk models (DeepSeek for Nigerian law) with EV justification

For the National Information Technology Development Agency (NITDA):

• Classify legal AI as "High-Risk" under any future AI regulation (following EU AI Act Article 6)
• Mandate transparency requirements: vendors must disclose training data composition by jurisdiction
• Establish testing and certification for legal AI tools using the methodology in this study
• Require incident reporting for AI-related legal errors

For Nigerian Law Schools (Nigerian Law School, university law faculties):

• Integrate AI literacy into curricula including:
• How LLMs work (stochastic parrots, Section 2.1 of literature)
• Documented failure cases (NG_009, NG_020 from Section 13)
• Verification protocols (Section 14.4)
• Domain-specific risk mapping (Appendix G)

For Legal Tech Vendors:

• Disclose training data composition and jurisdictional coverage
• Provide jurisdiction-specific accuracy data (as in Sections 3-6 of this report)
• Implement transparency features: confidence calibration displays, citation verification
• Support Nigerian-specific benchmarks (this study provides the template)
• Provide warnings about known failure modes (procedural blindness, document hierarchy)