BJ
Biography
In today’s dynamic digital landscape, artificial intelligence tools like large language model (LLM) based user interfaces popularly called “chatbots” have become indispensable for brainstorming ideas, drafting documents, and even strategizing legal defenses. However, a recent ruling from the Southern District of New York (SDNY) serves as a stark reminder that these innovative technologies can inadvertently jeopardize one of the cornerstones of the legal profession: attorney-client privilege. As specialists in AI and technology law at Buckley Law P.C., we regularly advise clients on how to harness emerging tech while safeguarding their confidential communications. This article explores the implications of this landmark decision, offering practical guidance to help you avoid potential pitfalls in your technology-driven legal practices.
Understanding Attorney-Client Privilege and Its Vulnerabilities with AI
Attorney-client privilege is a fundamental legal protection that shields confidential communications between a client and their lawyer from disclosure, ensuring open and honest dialogue essential for effective representation. This privilege applies only when communications are made in confidence, for the purpose of seeking or providing legal advice, and without waiver through disclosure to third parties.
The rise of AI chatbots—such as Anthropic’s Claude, Chat GPT, Gemini, Grok or similar generative tools—introduces new complexities. These platforms can analyze facts, generate arguments, and simulate legal strategies with remarkable efficiency. Yet, when users input sensitive information, they may unknowingly expose it to risks. Unlike secure, enterprise-grade systems designed for legal use, consumer AI tools often lack robust confidentiality safeguards. Their terms of service frequently allow data to be used for training models or shared with third parties, effectively turning the AI into an unintended “third party” that breaks the chain of privilege.
When users input sensitive information, they may unknowingly expose themselves to risks. Those confidential prompts are training data most publicly accessible and consumer grade chatbots. For the same or similar inputs, the chatbot may produce confidential names of parties, embarrassing facts, or your attorneys’ strategy enabling your adversaries to exploit that information given the right prompt. In one study, it took only five cycles for the chatbot to include data of the first query. How? Because user prompts entered into chatbots “train” the LLM to learn from user prompts. Dropping your attorneys’ memo into a chatbot to “get a second opinion” is a bell that cannot be unrung. Currently, innovation and creativity are for humans. Chatbots are slow in getting these skills right.