AI for Public Defenders

Step

What to Nail Down

Pro Tips

1. Security Review 

• Confirm vendor hosts data in CJIS‑compliant environment.
• Execute NDA & confidentiality addendum. 
• Decide what never goes into Gen‑AI prompts (e.g., client identifiers).

2. Pilot Design

• Choose one high‑volume use case (e.g., body‑cam transcription).
• Limit to 5‑10 attorneys & 50–100 cases.
• Set a start/stop date (e.g., 90 days).

3. Training Plan

• 2‑hour kickoff workshop: features, dos & don’ts, verification steps.
• “Cheat‑sheet” on prompt templates & red‑flag errors.

4. Success Metrics 

• Efficiency: avg. hours saved per case.
• Quality: % transcripts with ≤1 material error; motion‑grant rate.
• User adoption: log‑ins per week, survey of attorney satisfaction.

Tool Category

Specific Tools Discussed

Key Limitations

Primary Ethical Concerns

Mitigation Strategies/Requirements

Transcription (A/V Evidence)

JusticeText, Reduct.Video, LegalServer AI

Accuracy dependent on audio quality; potential for errors misrepresenting evidence 6

Competence (verifying accuracy); Confidentiality (handling sensitive recordings); Potential over-reliance due to time pressure 6

Mandatory human review & correction; User training on limitations; Vendor vetting for security/accuracy claims; Use human transcription for critical/court use 23

Legal Research & Drafting (GenAI)

Casetext CoCounsel, Westlaw AI, Lexis+ AI

Prone to “hallucinations” (false info/citations); Variable reliability; Requires skilled prompting 68

Competence (verification mandatory); Candor to Tribunal (avoiding false submissions); Confidentiality (inputting client data); Fees (billing for efficiency) 22

Rigorous verification against primary sources; Training on prompt engineering & limitations; Clear policies on data input; Client consent if needed; Fair billing practices 22

Case Management (AI Integration)

LegalServer AI Suite; (LACPD’s AWS/CMS integration)

Integration complexity; Accuracy of data extraction/classification; Potential for workflow disruption if flawed 45

Competence (understanding integrated features); Confidentiality (data within CMS); Supervision (staff use of AI features) 22

Thorough testing & validation of AI features; User training; Strong data governance within CMS; Human verification steps in workflow 23

Sentencing Analytics

SentencingStats

Reliance on historical data patterns; Predictive accuracy contested; Potential for misuse/overstatement 53

Bias (inheriting bias from historical data); Fairness (predicting future behavior); Transparency (algorithmic basis of prediction); Competence (interpreting stats) 6

Critical analysis of data sources & methods; Awareness of potential biases; Use as advocacy support, not definitive prediction; Transparency in methodology if possible 53

E-Discovery Platform (Broad AI)

Relativity (RelativityOne, aiR, Analytics)

Complexity; Cost/Access barrier; Requires skilled users/admin; AI feature reliability (e.g., aiR accuracy/cost) 83

Bias (in algorithms/data); Confidentiality (large scale ESI); Competence (using advanced features); Access Disparity (cost barrier for PDs) 6

Robust training; Skilled personnel; Vendor vetting (security, AI validation); Clear protocols for AI feature use; Advocacy for resources for PD access 23