AI ETHICS The Vanderbilt white paper states that “decision-making in charging, sentence recommendations, and other outcomes” could be “the most consequential use of AI powered systems in prosecution.” 71 The white paper itself notes that “there is a major risk in lack of validation, and by extension, the potential for perpetuating biases” should AI replace prosecutorial judgment about case screening, charging, bail recommendations, plea negotiations, sentencing recommendations, and post-conviction assessments. 72 It is essential for defenders to anticipate even minimal AI use by prosecutors for these high-risk tasks, and defenders should likewise consider their use of AI for these tasks as high-risk. Of course, there are some lower-risk tasks contained in those high-risk tasks that prosecutors are already using AI for, and defenders should do the same. Those include using AI (LLMs or legal-specific tools) to summarize case files. One recommendation for using AI in these specific instances would be to have an “independent researcher review a random sample of decisions and determine whether [] disparities exist in the outcomes produced.” 73
4.
Confidentiality Protocols
Part II’s discussion of Rule 1.6 established the confidentiality framework. This section translates that framework into protocols.
The basic rule is that defenders should use only secure, approved systems for work involving confidential information. What constitutes “secure” and “approved” requires office-level determination based on available resources and risk tolerance, but certain minimum standards apply universally. Any system handling client data should operate in a closed environment, either as a cloud-based instance or an on-prem environment. The system should encrypt data in transit and at rest. Access should be limited to authorized personnel with legitimate need. And the vendor should contractually commit to confidentiality protections at least as strong as the defender’s own obligations. Defenders should establish clear categories of information and corresponding protocols for other types of non-confidential information. Some information — such as general legal research unconnected to any client — may appropriately be processed through less secure systems, including open-sourced and open environment systems like a free model of ChatGPT or Claude. Other information — including case facts, client communications, and discovery materials — should never enter any AI system without enterprise-grade security. The most sensitive information — details that could endanger witnesses, privileged strategy discussions, or materials subject to restrictive protective orders — may warrant exclusion from AI systems entirely, even secure ones. Practical implementation requires training and enforcement. Staff must understand which systems are approved for which purposes. Workflow design should channel work toward appropriate tools and make inappropriate uses inconvenient. Regular audits can verify compliance. And a culture that supports reviewing each other’s AI practices also reinforces good habits.
37
Parity in Practice: The Defender’s Duty to Ethically Use AI
Made with FlippingBook Online document maker