AI-generated large language models (LLMs) are gradually entering every aspect of our lives, including court decisions. Judges now have greater clarity on when it is appropriate or inappropriate to rely on these tools due to new guidelines that the UK’s Judicial Office released this week. Judges are advised not to use the tools for creating new analyses by the UK guidance. It does, however, permit text summarization. Meanwhile, a growing number of US lawyers and defendants are facing fines and penalties for carelessly incorporating artificial intelligence (AI) into their legal practices.
The Judicial Office’s AI guidance is a collection of ideas and recommendations meant to assist judges and their clerks in comprehending artificial intelligence (AI) and its limitations as the technology becomes more widely used. These rules are not legally binding; rather, they represent the “first step” in a series of initiatives by the Judicial Office to make clear how judges can use technology.
According to the new guidelines, judges may find ChatGPT from OpenAI and other AI tools helpful for administrative tasks like drafting memos or emails or as a research tool for summarising large amounts of text. Judges were cautioned against using instruments for legal research that depend on fresh data that cannot be independently verified at the same time. The guidelines simply state that public AI chatbots “do not produce convincing analyses or reasoning” when it comes to crafting legal arguments. Judges should avoid using the tools to conduct new research on subjects they are unable to independently verify, but they may find some advantages in using an AI chatbot to retrieve information from the guidance notes that they already know to be accurate. It seems that the guidance places the responsibility of differentiating fact from fiction in the LLM outputs on the user.
The advice states, “It may be best to view them [AI tools] as a means of obtaining non-definitive confirmation of something, rather than providing immediately correct facts.”
The guidelines go on to caution judges that even when given extremely specific or precise instructions, artificial intelligence (AI) tools may nevertheless produce biased, erroneous, or incomplete results. Most people refer to these strange AI fabrications as “hallucinations.” Similarly, because a number of the services are “open in nature,” judges are cautioned against entering any “private or confidential information” into the system.
The guidelines state that “any information that you input into a public AI chatbot should be seen as being published to all of the world.”
It is unclear exactly what use “private” information fed into the LLM would serve in the legal context, given that the information spat out from a prompt is “non-definitive” and possibly inaccurate and that information fed into the LLM cannot include “private” information that might be crucial to a thorough review of, say, a lawsuit’s text.
The Judicial Office is also concerned with context-dependent data. The most well-known AI chatbots currently available on the market, such as Google’s Bard and OpenAI’s ChatGPT, were created in the US using a sizable corpus of data with a US focus. The recommendations caution against giving AI models a “view” of the law that is biassed towards American legal contexts and theory if they focus primarily on US training data. Ultimately, as the guidance points out, judges are ultimately accountable for content created in their names, even if an AI tool was used to assist in the process.
Prior to the release of the guidelines, Geoffrey Vos, the Head of Civil Justice for England and Wales, reportedly told that he thought AI would present significant opportunities for the justice system. He continued by saying that he thought judges could identify arguments made with artificial intelligence.
Judges are trained to distinguish between the true and the false, and Vos stated that they will need to do so in the AI-driven modern world just as much as they did in the past.
Despite accuracy concerns, some judges already find AI to be “jolly useful.”
The updated guidelines were released three months after Lord Justice Birss, a judge in the UK court of appeals, utilized ChatGPT to summarise a legal topic before using some of the information to draft a decision. At a press conference earlier this year, Birss stated that even though the judgement was written with the assistance of an AI tool, he should still be held ultimately responsible for its content.
Birss said that, he is not trying to give the responsibility to somebody else; he is taking full personal responsibility for what he put in his judgement. It accomplished nothing more than set him up for a task for which he was already prepared, knew the solution, and could accept.
Some lawyers and defendants have already found themselves in hot water due to a lack of clear regulations defining when and how AI tools can be used in court documents. A couple of US lawyers received a $5,000 fine earlier this year for submitting a court filing that included fictitious citations produced by ChatGPT. A UK woman was allegedly caught using an AI chatbot to defend herself in a tax case more recently. After it was found that the case law she had submitted contained fabricated details that the AI model had hallucinated, she ultimately lost her appeal. Earlier this year, ChatGPT was accused of authoritatively naming a radio show host as the defendant in an embezzlement case that he had nothing to do with, leading to a libel suit against OpenAI.
It’s possible that the ambiguity surrounding AI in legal proceedings will worsen before improving. Congress hasn’t been able to pass any comprehensive legislation that sets clear guidelines, despite the Biden Administration’s recent AI Executive Order offering proposals governing the deployment of AI in legal settings. Conversely, the European Union has recently reached a consensus on its AI Act, imposing more stringent safety and transparency regulations on a broad spectrum of AI instruments and uses that are classified as “high risk.” However, it’s unlikely that the actual sanctions for breaking those regulations will be implemented until at least 2025. Juries and attorneys are therefore currently essentially winging it when it comes to determining the moral bounds of AI use.
