How can lawyers stop AI's hallucinations? More AI, of course.

In a bid to combat the issue of AI-generated inaccuracies in legal documents, law firm Cozen O'Connor is implementing an advanced AI hallucination detection system. This decision comes in light of a recent incident where two of its lawyers faced penalties for citing fictitious cases in their filings. The firm is now collaborating with a startup named Clearbrief, which has developed software capable of scanning legal briefs to identify fabricated facts and citations, functioning similarly to spell check but focused on legal accuracy. Kristina Bakardjiev, a partner at Cozen O'Connor, emphasized the necessity of adapting to technological advancements, acknowledging that lawyers are likely to experiment with chatbots regardless of established policies. The legal industry has reacted to the risks posed by AI hallucinations by instituting bans on general-use chatbots, but the challenge remains in preventing associates from utilizing these tools inappropriately. In September, the two defense attorneys from Cozen O'Connor admitted to submitting a document filled with false citations after using ChatGPT, which violated firm policy. The situation escalated to the point where a Nevada district court judge imposed a $2,500 sanction per lawyer, offering them the option to either withdraw from the case or publicly address their misconduct at their former law schools. Damien Charlotin, a legal data analyst, has been monitoring cases where courts have identified AI-generated inaccuracies, noting a rapid increase in reported incidents—from 120 cases between April and May 2023 to 660 by December, with new occurrences rising to four or five daily. While these numbers are relatively small in the grand scope of legal filings, they predominantly involve self-represented litigants or smaller firms, with larger firms often facing issues through less experienced staff. The challenges posed by AI hallucinations extend beyond the legal field. In recent months, consulting firm Deloitte had to issue a partial refund to the Australian government for a report found to contain numerous AI-generated inaccuracies. AI hallucinations are deeply ingrained in the functioning of large language models, which predict subsequent words based on previous inputs. Although complete elimination of these inaccuracies is unattainable, experts suggest that companies can significantly mitigate risks by limiting AI models to specific datasets, such as curated legal databases. Legal research giants Thomson Reuters and LexisNexis are now promoting AI tools designed to assist lawyers in efficiently navigating and citing legal data. These firms have invested substantial resources in building extensive repositories of legal content, which they believe offer safer alternatives to open internet-based chatbots. Clearbrief, the company behind the detection software, has created a tool that integrates with Microsoft Word, helping litigators ensure the accuracy of their citations and facts. This innovative approach not only flags errors but also provides direct links to the relevant legal documents, enhancing accountability in legal drafting. Cozen O'Connor is also upgrading its knowledge management system to include citation check reports, creating a comprehensive record of document accuracy that can be referenced if questioned by a judge. As the legal profession grapples with the prevalence of AI hallucinations, training lawyers to view chatbot outputs as preliminary resources rather than final products will be essential for maintaining integrity in legal practices.