Building Safeguards Around Artificial Intelligence

This is the third of six articles in our series on GenAI use in practice.  We thank our contributing counsel for their responses: Denise Brunsdon (DB), Codie Chisholm (CC), Adam Drew, KC (AD), Katrina Edgerton-McGhan (KEM) Jennifer Hawkins (JH), Jivanya Reddy (JR), Ryan Robertson (RR), Marcus Sixta (MS), Dylan Snowdon (DS), Laura Warner (LW), and Jenna Weinkauf (JW).

Law Matters asked: “How have you applied safeguards to ensure accuracy and mitigate risks in using GenAI tools?"

AD: We have a detailed protocol, and closed spaces to work in. I follow these safeguards carefully. I also try to be very skeptical, asking myself “Is this something I would push back against if it was recommended by a human?” “Does this output still match my voice and intended message?” “Am I sure that all the documents actually say what the AI suggests, and can I confirm that myself?”

CC: I follow my firm’s AI policy, and I am one of the lead educators at my firm on responsible AI use in the legal profession. Our firm has invested in enterprise AI models that protect clients’ confidential and privileged information. For legal research, I run numerous prompts to identify cases or other authorities relevant to the questions I am researching. I always verify the output by reviewing the source document.

DB: Safeguards are essential, and we have many. The recent headlines about case hallucinations reinforce why lawyers must approach these tools carefully and maintain their professional obligations. Lawyers have always been responsible for verifying facts and law and protecting privileged information, and AI does not change that. Using secure, firm-approved tools, verifying outputs, and being transparent about AI use are key aspects of responsible adoption. We also have an obligation to embrace responsible AI use for the benefit of training new lawyers, and hopefully, increasing access to justice overall.

DS: My firm has implemented policies and procedures to ensure that staff and lawyers are only using approved tools that ensure confidentiality and privilege are maintained, and only for explicitly approved tasks. We have also ensured that all lawyers and staff are educated about the risks different AI tools bring, including potential breaches of confidentiality and inaccurate information.

JW: Our firm uses closed AI tools with appropriate confidentiality safeguards.  We delete projects from the AI tool after the matter is completed so that there is no lingering information in the ether.  We also take steps to ensure the AI does not supplant legal judgment.  On my own files, I review AI’s work and tailor the output to match what is required for the matter.  Across the firm, work prepared or edited by AI must be reviewed by a qualified lawyer before being sent to a client for comment.

JH: Our firm purchased a closed AI program that encrypts information as it goes in and out of the program. The closed AI program does not learn from our work and all information is housed in Canada. As a further safeguard, our lawyers anonymize notes before they are inputted into the closed AI program.

Our retainer letter clearly outlines how we use AI and clients are asked to advise if they do not want us to use it.

Summaries of evidence generated by our AI program are carefully reviewed by the investigating lawyer, and then, as for all our reports, by our editor and a senior lawyer.

JR: I treat anything generated by AI as a draft, not as an authority. Legal information must still be checked against legislation, case law, and other reliable sources. 

KEM: We have a closed AI program that complies with all Canadian privacy laws.  I treat anything generated as a first draft and check each point with the primary source for reliability.

LW: I control the information that AI relies on. I closely check its outputs for reliability. I ask it to play devil’s advocate.

MS: We have included terms on AI use in service agreements. This is not just because we may use AI directly from time to time, but also because many common programs now have AI integrated as a default. Also, before using AI tools directly on a client’s file, we seek the client’s written consent. In addition, any time we use an AI notetaking device we seek consent, as AI notetakers record conversations and we understand our ethical duty to seek consent before recording a conversation.

We also understand that any AI generated content must be reviewed for inaccuracy and hallucinations. There is a direct link between an AI's drive to provide a helpful or positive response and its tendency to hallucinate. This has led to some high-profile cases where lawyers have been disciplined for relying on AI hallucinated law in court.

RR: We treat AI output as a draft or a prompt for further review.  A lawyer checks substantive suggestions against the underlying documents, facts, and applicable laws before using them.  We also assess the confidentiality and security of a tool before putting client information into it.