Attorneys are adopting artificial intelligence faster than they are establishing the guardrails to use it safely. Jennifer L. Ellis, an attorney and consultant who advises lawyers, law firms, and legal organizations on ethics, technology, and AI, has been closely tracking the consequences.
We spoke with her about the mistakes that can cost attorneys their credibility — and how to avoid them.
Our founder and CEO, Shannon Wilkinson, met Jennifer when she presented to the New York Regional Expert Witness Association. Shannon serves as an expert defamation witness and regularly works with attorneys.
Q: AI has become almost unavoidable in legal practice. From your perspective as an ethics attorney, what is the primary reputation risk attorneys face when they adopt AI tools without proper guardrails?
A. Currently, the major risk is using hallucinated cases in court documents and being publicly called out and/or sanctioned for failing to read and verify cases. An attorney who fails to read and verify the cases violates quite a few ethics rules, but more, these cases frequently get a lot of attention. Many lawyers write blog posts or create videos about the cases which spread the potential for damage. Some lawyers have faced or are facing sanctions from the disciplinary body of their jurisdiction(s). Public sanctions never go away, and if there are many articles and/or blog posts when someone searches the name of the lawyer, such articles may be surfaced while potential clients are researching whether to hire them.
Q: You’ve developed what you call a “Ladder of Risk for AI Tasks,” a framework to help attorneys evaluate AI use before getting started. Can you provide an example of a low-risk and a high-risk use?
A. First, I should mention that there is always a risk with Generative Artificial Intelligence (GAI). There is no GAI tool that does not present a risk of hallucination, or what I will call disobedience by the AI. As far as specifics, a low-risk task is one that is entirely internal and will not leave your office or be seen by anyone else. For example, taking disorganized notes and turning them into something else. Somewhat low risk would be summaries of internal documents, but the AI could leave out something important or add something that it made up. High risk includes anything being filed with a court.
Q: Hallucinated citations — when AI tools generate cases that don’t exist — have already resulted in sanctions. How significant is this reputational risk, and how should attorneys protect themselves?
A. The risk is reputational and the best way to avoid it is to verify everything. This means if you use AI for legal research, you should find and read the case. First to make sure the case exists, second to make sure it stands for the proposition for which it was cited. More judges are questioning the non-filing lawyer as well. This could include a supervisory lawyer, who must properly supervise attorneys under them, or opposing counsel that fails to catch and report the problem.
Q: Many attorneys assume that using AI for administrative tasks — such as scheduling, billing, or formatting — is essentially risk-free. Is that accurate?
A. It is not. GAI is always at risk of hallucinations, fabrications, and disobeying instructions. For example, I instructed Claude to never use em dashes. In response, at first Claude started using two dashes. I asked if it was doing so to get around my em dash instruction. It responded that it was sorry, it knew it shouldn’t have done it, but yes. Sometimes it still uses em dashes. This is because em dashes are so baked into its training that it simply cannot avoid using them. Disobey isn’t a great word because it suggests intent, and AI doesn’t intend to fail to follow my instructions, but it is a good word for helping users understand the problem.
This risk is even greater with agentic AI. Agentic AI can complete many steps all by itself. The problem is if it hallucinates or disobeys at some point in the process, this can cause a risk of cascading failure. For example, one company had its agentic AI delete its entire database and backups in 9 seconds.
Q: Confidentiality is a foundational duty in law. What risks do attorneys take when entering client information into publicly available AI tools, and how common is this practice?
A. I suspect the problem is extremely common, not just among attorneys but among clients, which creates different problems. For clients, the use of AI can lead to loss of privilege and discovery of everything they enter into the AI for litigation. For lawyers, using AI in consumer-level tools without the proper privacy options would be a violation of Ethics Rule 1.6, confidentiality. It would be hard to know if an attorney actually used a tool they should not at this point, so to a certain extent it is on the integrity of the attorney to protect their client’s privacy by checking the privacy policy and terms of service as well as properly anonymizing confidential information if they use a tool that does not meet the proper privacy requirements.
Q: You recommend using enterprise-grade tools rather than consumer AI products when handling client materials. What distinguishes the two, and how should attorneys evaluate whether a tool is trustworthy?
A. What distinguishes the two is the ability to control what happens to the data, who can access it, whether it can be deleted immediately, and whether the confidentiality is protected by contract. There are business-level tools which may be acceptable as well. Claude has a tool that offers greater protection than consumer grade if you purchase 5 seats for $125/month. ChatGPT has one that has greater protection too. You purchase at least 2 seats for $50/month. You still need to check the privacy policy and the terms of service to make sure that the tools are acceptable under confidentiality rules. You also need to be concerned about laws that protect certain types of data, such as HIPAA for certain medical information.
Q: Some attorneys worry that being cautious with AI will put them at a competitive disadvantage. What is your advice to those who feel pressure to adopt AI more aggressively?
A. The two are not mutually exclusive. I use consumer-level tools frequently. I simply make sure to properly redact any confidential data. It would be more difficult to redact data such as thousands of pages of medical records, which means that in that case, it is critical to use a tool that has the proper level of protection. The fear of being put at a competitive disadvantage is not unreasonable though. Especially when it is data that is hard to redact, or a court puts a protective order in place which requires use of tools that are at the enterprise-level. The cost of such tools is frequently too great for a solo or small firm, or an individual who is representing themselves pro se.
Jennifer L. Ellis is a Pennsylvania attorney and legal technology consultant focused on ethics, law practice management, and the practical use of technology in legal settings. She advises lawyers and legal organizations on technology selection and implementation, content strategy, professional responsibility issues, and continuing legal education. She frequently speaks and writes on legal ethics, cybersecurity, and emerging technologies, including AI.
Recommended reading: her blog post, The AI Ladders of Risk.
*
This is part of our continuing series of interviews with experts whose work relates to reputation management.
