AI ETHICS
c.
Why does this matter?
For defense lawyers, this distinction matters enormously. Entering privileged client information, case facts, discovery, or any other materials subject to protective orders into either system risks breaching confidentiality obligations under Model Rule 1.6, but closed systems remain a safer option for defenders who disclose this type of information in an AI tool. State bar opinions have likewise required lawyers to understand how different AI tools handle data. 32 Even if the risk of a specific query resurfacing in another user’s results seems remote, the possibility is not zero and defenders must weigh that risk against the potential benefits.
3.
Examples in Practice
Several AI tools have emerged as relevant to legal practice. General-purpose tools include ChatGPT (developed by OpenAI), Claude (developed by Anthropic), and Gemini (developed by Google). These LLMs allow users to generate text (e.g., drafting an email to the client or preparing a section of a brief or motion), summarize documents, and receive answers to questions across virtually any subject matter. Legal-specific AI tools have also proliferated. Thomson Reuters has integrated AI capabilities into Westlaw through its CoCounsel product, which was the first legal AI assistant powered by OpenAI’s GPT-4 technology and now provides agentic functionality using Anthropic’s Claude SDK (software development kit). LexisNexis offers Lexis+ AI for legal research and document analysis. Harvey, a startup backed by Sequoia and OpenAI, has partnered with major law firms to develop AI tools tailored specifically for legal work. But legal-specific AI tools share the same problems as other AI tools and LLMs. Legal tech startups often promote “retrieval-augmented generation” (RAG), which act as LLMs but instead of drawing from the entire internet, RAG models “get their information only from a closed set of data, such as Westlaw, Bloomberg, or the lawyer’s own collection of documents, allowing RAG systems to provide accurate citations when answering user prompts.” 33 Although legal AI tools may reduce errors compared to the general-use LLMs, RAGs still hallucinate “at an alarming rate.” 34 Researchers at the Stanford RegLab and Institute for Human- Centered Artificial Intelligence (HAI) found “the Lexis+ AI and Ask Practical Law AI systems produced incorrect information more than 17% of the time, while Westlaw’s AI-Assisted Research hallucinated more than 34% of the time.” 35 Defenders evaluating these tools should consider several factors, including but not limited to: (1) whether the system operates as open or closed (both source and environment); (2) what data security guarantees the vendor provides, such as data retention and deletion controls, encryption standards, role-based access controls, and breach notification procedures; (3) how the system was trained and on what data; and (4) what verification mechanisms exist to catch errors or hallucinations in the output.
15
Parity in Practice: The Defender’s Duty to Ethically Use AI
Made with FlippingBook Online document maker