AI ETHICS
5.
Accuracy and Verification Protocols
Written verification expectations establish minimum standards and create accountability. The policy should require attorney verification of all AI outputs before submission to courts, specify that citations must be checked against official sources (with a specific callout that many legal- specific AI tools still hallucinate), establish documentation requirements for verification processes, and define what “verification” means for different types of outputs. For example, verification expectations may be less burdensome for general legal research conducted by a junior attorney to learn more about a particular legal issue common to the firm or office, but more burdensome when using AI to make strategic decisions about a client’s case or to draft a motion or brief that will be submitted to the court. The strongest verification expectations should correlate to high-risk tasks, such as the submission of evidence or exhibits that were created or modified by AI. The policy might distinguish between verification requirements for different risk levels. High-stakes submissions — like sentencing memos — should require documented verification by at least one, if not multiple, human reviewers. Lower- stakes documents might permit less elaborate verification while still requiring human review.
6.
Supervision and Training
Effective AI use requires training, and compliance requires supervision. The policy should mandate initial training for all personnel before they use AI tools, require periodic refresher training as tools and policies evolve, establish supervisory review requirements for AI-assisted work product, and create mechanisms for answering questions and addressing concerns. Training content can be developed internally or provided by the various AI vendors (some vendors provide robust training and onboarding support). 77 The content should cover the office’s AI policies and procedures, the capabilities and limitations of approved tools, confidentiality and verification requirements, recognition of AI errors and hallucinations, and compliance with court orders regarding AI disclosure. Training should be practical, including hands-on exercises with actual tools rather than instruction alone, when practical. This training could include how to effectively and efficiently generate prompts, reduce wasteful use, and otherwise use a specific tool, or how to spot hallucinations, overstatements of law, and other pitfalls of AI-generated content.
7.
Vendor and Procurement Standards
Policies should establish criteria for evaluating AI vendors and guide procurement decisions. Required vendor characteristics might include demonstrated security practices and certifications, contractual commitments to confidentiality and data protection, transparency about data handling and retention, financial stability and business continuity planning, and responsiveness to security inquiries and incident notification. Others might include sustainability efforts, training and onboarding offerings,
43
Parity in Practice: The Defender’s Duty to Ethically Use AI
Made with FlippingBook Online document maker