Since the publication of my technical article “My ‘Hallucinating Experience with ChatGPT’”, I have struggled with the extent to which I should or should not be using artificial intelligence (AI) to perform my judicial duties. It seems that other judicial officers have engaged in similar soul-searching. Notwithstanding these internal debates, we should not overlook that our profession has a history of using AI tool, starting with spell check in the 1980s and, subsequently, the advent of predictive typing and grammar-checking computer applications. Later, computer-aided legal research became a big thing we enthusiastically embraced. But now, AI!
Many judges have begun a personal journey to understand the contours of AI’s capabilities for their judicial duties. Most of the early articles I read on the subject included:
1) warnings about the limitations of AI,
2) concerns about whether AI tools are operating in a fair and nonbiased manner, and
3) stringent reminders that each judge is solely responsible for their judicial decisions and cannot rely solely on AI.
It is now gratifying to see that courts and other entities have started to study the risks and benefits of AI to the judicial profession and provide guidance to judicial officers regarding their use of AI.
Among the entities that have produced recommendations or guidance to judges are the National Center for State Courts (NCSC) in its publication “Artificial Intelligence: Guidance for Use of AI and Generative AI in Courts”; the American Association for the Advancement of Science in a group of publications entitled “Artificial Intelligence and the Courts: Materials for Judges”; the Illinois Supreme Court Policy on Artificial Intelligence; and the New Jersey Courts Judiciary Policies and Guidance on AI. These efforts are the beginning of many similar efforts. Indeed, other courts and entities have announced or formed committees or working groups to study the risks and benefits of AI to the judiciary. These multiple efforts will only enhance the ability of judges in our numerous judiciaries to understand the risks and benefits related to judicial use of AI and exercise competence in exploiting AI’s vast potential.
In the interim, however, there are a few basic traps that everyone is likely to encounter when using AI, which I will highlight with a few plausible examples of judicial and nonjudicial use of AI. Readers are cautioned that these examples barely scratch the surface of the numerous considerations judges must undertake to pursue an informed use of AI technology in carrying out their judicial duties.
Starting with a nonjudicial use of AI, assume I am writing a short speech to give at a friend’s retirement. The specific AI tool used makes little difference. It could be ChatGPT, Bard, CoPilot, Gemini, or another popular AI tool. The speech I drafted is okay but not exactly what I want. I put the speech into my AI tool with a prompt to revise it for clarity and make it thoughtful and humorous. In response, the AI tool provides a new version of the draft speech with several revisions. Two sentences of the suggested revisions are fantastic, and I use them fully without attribution to my AI tool. Several AI suggestions were new thoughts that helped me refine my initial draft. I have decided to reject the remaining suggestions by the AI tool because I question whether they are better than what I originally wrote. Here are some problems that are not readily apparent to someone in their beginning efforts with AI.
Security, Data Privacy, and Confidentiality
The speech included my friend’s name, his employer, and facts about his life and our decades-long relationship. All that information is now included among the terabytes of data in that AI system for future unknown uses. Indeed, someone might make a future inquiry to the same AI system about my friend or his company and receive a response that includes some of the information I provided in my speech. Suppose I included secrets about my friend in my initial drafts that I later decided not to use because they were too personal or embarrassing. That information is now in the AI’s databank and may be revealed. It can potentially be discovered by others using the same AI tool.
Copyright Infringement and Plagiarism
Another problem that arises from this exercise is the absence of any disclosure by the AI tool concerning the sources used to develop the suggested revisions. Did they come from copyrighted materials? Were they a plagiarized version of someone’s blog post or online dissertation? Were the suggestions taken from someone else’s plagiarism? Am I obligated to attribute the remarks to some source?
The same concerns arising from using AI to improve my speech also exist if I, as a judge, write an order that includes sensitive and personal information about the parties and use AI to obtain suggested revisions to achieve clarity and more thoughtfulness and find relevant legal citations. The information I have now submitted to the AI system can be spread to others far beyond what might typically occur with an ordinary court order. If the AI tool suggests revisions that include citations to legal authority, are those references accurate, nuanced, or even imaginary? Who among us has not heard about lawyers using AI to write legal memoranda that came with fake legal citations? Suppose the case involved health data or other privileged or nonpublic information and the AI tool does not prioritize confidentiality and security. In that case, the submitted information is now in the data bank of the AI tool for future unrestricted use. Further, as a judge, am I aware of whether my law clerks and assistants use AI tools when preparing drafts for my consideration? Are they submitting this information to some AI tool that will have unfettered future use of the information they submitted?
Lastly, suppose I use my AI tool to summarize the voluminous pleading and exhibits I received in a pending case. I will likely receive an impressive summary of each party’s arguments and supporting documents. There are several questions to ask: To what extent is the AI response an accurate summary? Were some arguments weighed more heavily than others? Is the summary misleading because the AI’s algorithms are such that certain issues are emphasized more than others? Was any part of the AI summary misleading, a fabrication, or, using the euphemistic language of the AI creators, a hallucination? How different is this scenario from a summary provided by a law clerk or assistant? How familiar am I with the materials being summarized? These are all essential considerations for a judicial officer using AI to perform judicial duties.
Considering these limited examples, we begin to see some of the concerns raised by judicial use of a particular type of artificial intelligence, GenAI (Generative AI), a form of artificial intelligence that can generate new content (including text, images, sound, or video) in response to a prompt.
AI tools are designed to learn from an expansive reference database. From the prompts above, the AI tool responded by searching its database for similar examples of the content being sought. For instance, ChatGPT was trained on a dataset of 300 billion words. This massive database was cumulated from numerous sources, including web searches consisting of web crawls over eight years, Reddit posts with three or more upvotes, internet-based fiction and nonfiction book collections, academic articles, Wikipedia, and more. In addition, ChatGPT responses to prompts are controlled by over 100 trillion parameters—factors that form or define the system or set the conditions of its operation for responding to prompts.
Institutional Development of AI Policies
The NCSC report encourages court institutions to develop internal AI use policies. According to the NCSC’s guidance, courts must be aware of the capability and limitations of GenAI tools, including issues related to accuracy, bias, ethics, and security. By extension, the individual judges of those courts also must be aware of those issues.
The report states that court leaders should establish policies that enable their organizations to experiment with and benefit from AI technologies while at the same time minimizing risk. The report urges courts to establish working groups to oversee the acceptable use, development, and management of AI technologies and policies and assess the risks associated with implementing an AI tool in areas such as hallucinations, data security, bias, and copyright infringement. Finally, in urging courts to develop AI use policies, the report notes the policies should include:
1) to whom and for what technologies the policies apply;
2) the acceptable and prohibited uses of AI;
3) a listing and explanation of the applicable data protection laws, regulations, and security measures that should be implemented;
4) how to ensure that AI-generated content is not biased and does not reflect illegal discrimination;
5) when to update and patch AI tools to protect against vulnerabilities and security risks; and
6) mechanisms to monitor and enforce the policies.
Defining the proper and efficient use of today’s rapidly changing AI technology should not be left to a few self-motivated “techie” judges. Active institutional involvement, study, and development of policy are essential.
Ethical Concerns
Today’s American Bar Association (ABA) Model Code of Judicial Conduct, adopted by the ABA in 1924 as the Canons of Judicial Ethics, is approaching its first anniversary. As discussed by W. Kearse McGill, a judge for the California State Bar Court, the nation’s only trial and appellate court system dedicated solely to adjudicating attorney discipline cases, there are several ethical rules a judge should consider when using generative artificial intelligence, as follows:
- Canon 1, Rule 1.2: “A judge shall act . . . in a manner that promotes public confidence in the independence, integrity, and impartiality of the judiciary. . . .”
- Canon 2, Rule 2.2: “A judge shall uphold and apply the law and shall perform all duties of judicial office fairly and impartially.”
- Canon 2, Rule 2.3(A): “A judge shall perform the duties of judicial office . . . without bias or prejudice.”
- Canon 2, Rule 2.4(B): “A judge shall not permit family, social, political, financial, or other interests of relationships to influence the judge’s judicial conduct or judgment.” Comment [1]: “Confidence in the judiciary is eroded if judicial decision making is perceived to be subject to inappropriate outside influences.”
- Canon 2, Rule 2.5(A): “A judge shall perform judicial and administrative duties, competently and diligently.” Comment [1]: “Competence in the performance of judicial duties requires the legal knowledge, skill, thoroughness, and preparation reasonably necessary to perform a judge’s responsibilities of judicial office.”
- Canon 2, Rule 2.7: “A judge shall hear and decide matters assigned to the judge. . . .”
These rules are implicated by several concerns, including:
1) the erosion of public confidence in the judiciary if the public perceives that a judge relied on AI tools to decide cases, as opposed to the judge’s experience, education, and skill;
2) AI bias that might affect a judge’s decision; and
3) a judge’s competence understanding AI technology, especially GenAI, and its risks and limitations. While there has been debate about whether Comment [1] to Rule 2.4(B) requires the technological competence of judges, that issue is now in the past.
Two jurisdictions have specifically interpreted their state rules to require judicial technological competency. The State Bar of Michigan issued a 2023 opinion that judicial officers are ethically obligated to maintain competence with and educate themselves on advancing technology, including artificial intelligence. Similarly, the West Virginia Judicial Investigation Commission issued an advisory opinion in 2023 affirming a judicial ethical obligation for competence in technology, advising judges to use AI with caution, to supervise its outputs as if a human law clerk had produced them, and never to use artificial intelligence to decide the outcome of a case. The opinion states: “Judges should think of AI as a law clerk, who is often responsible for doing a judge’s research. Importantly, the law clerk never decides the case. The judge alone is responsible for determining the outcome of all proceedings.”
Final Thoughts
The first takeaway from this technology column is that courts, as institutions, must step up to guide their judicial officers on the use of AI technology by establishing working groups, studying the benefits and risks to judges of using various AI technology tools, and setting policies to guide the judges and their use of AI technology in the performance of their judicial functions.
The second takeaway concerns accountability. As I stated in a previous writing, when you are presented with a draft document prepared by an AI-powered product, as with a draft document prepared by a law clerk, paralegal, or new lawyer whose abilities you have not fully assessed, it is up to you to exercise due diligence before you sign the document. The West Virginia advisory opinion takes the warning further, adding, “. . . [the judge] cannot say, ‘the law clerk made me do it’ [or] ‘AI made me do it.’ The responsibility for the finished product rests solely with the judge.”