The Alaska Court System's project to launch an artificial intelligence chatbot to assist residents with probate matters has faced significant delays, extending a planned three-month initiative to over a year. The core challenge has been ensuring the AI's complete accuracy in a high-stakes legal environment where incorrect information could cause serious harm to families.
The project, known as the Alaska Virtual Assistant (AVA), was designed to simplify the complex judicial process of transferring a deceased person's property. However, issues with AI-generated falsehoods and the intensive human oversight required have highlighted the practical difficulties of deploying generative AI in sensitive government services.
Key Takeaways
- An AI chatbot project for Alaska's court system, intended to take three months, has now extended beyond 15 months.
- The primary obstacle is the AI's tendency to produce inaccurate information, a risk deemed unacceptable for legal guidance.
- Developers had to fine-tune the AI's personality, removing empathetic responses that users found unhelpful.
- The experience demonstrates the significant gap between AI hype and the labor-intensive reality of implementing it safely in the public sector.
An Ambitious Goal Meets a Difficult Reality
The Alaska Court System initiated the AVA project with the goal of creating a low-cost, digital equivalent to its human-staffed family law helpline. The vision was to provide residents with a reliable tool to navigate the often confusing forms and procedures of probate, especially during a difficult time of loss.
Stacey Marz, the administrative director of the Alaska Court System, explained that the standard for this technology had to be exceptionally high. Unlike other tech rollouts where a "minimum viable product" might suffice, the potential consequences of flawed legal information from an AI are severe.
"If people are going to take the information they get from their prompt and they’re going to act on it and it’s not accurate or not complete, they really could suffer harm," Marz stated. "It could be incredibly damaging to that person, family or estate."
This commitment to 100% accuracy has been the main driver behind the project's extended timeline. What was conceived as a quick technological leap has turned into a meticulous, year-long process of testing and refinement.
The Broader Context of AI in Government
The challenges faced by the AVA project are not unique. A recent Deloitte report found that less than 6% of local government practitioners were prioritizing AI as a tool for delivering services. The Alaska experience illustrates why: concerns about reliability, trustworthiness, and the need for constant human oversight create significant barriers to adoption, especially in high-stakes fields like the justice system.
The Challenge of AI Hallucinations and Personality
A significant technical hurdle for the development team was the phenomenon of AI "hallucinations," where the system confidently presents false information as fact. Aubrie Souza, a consultant with the National Center for State Courts (NCSC) involved in the project, recounted a specific instance of this problem.
"We had trouble with hallucinations... where the chatbot was not supposed to actually use anything outside of its knowledge base," Souza said. She explained that when asked for legal help resources, the AI invented a non-existent law school in Alaska and suggested contacting its alumni network.
To combat this, the chatbot's developer, Tom Martin of LawDroid, has worked to restrict AVA to a closed knowledge base consisting only of the Alaska Court System's official probate documents. This prevents it from searching the wider internet and fabricating answers.
Finding the Right Tone
Beyond factual accuracy, the team also had to carefully calibrate the AI's personality. Early versions of AVA were programmed to be empathetic, offering condolences to users who were likely grieving. However, user testing revealed a surprising reaction.
"Through our user testing, everyone said, ‘I’m tired of everybody in my life telling me that they’re sorry for my loss,’" Souza noted. The feedback indicated that condolences from an AI felt insincere and unhelpful. As a result, these empathetic expressions were removed to create a more direct and functional user experience.
The Cost-Benefit Analysis
Despite the development challenges, cost remains a compelling reason for government agencies to explore AI. The AVA team sees the technology as a potential solution for courts with limited budgets. According to developer Tom Martin, one technical configuration could process 20 user queries for just 11 cents, showcasing the potential for long-term financial efficiency.
A Labor-Intensive Process of Verification
The project underscores that implementing AI is far from a hands-off process. The team initially designed a comprehensive test with 91 questions to evaluate AVA's accuracy. However, the process of running the test and having legal experts manually verify each response proved too time-consuming.
According to Jeannie Sato, the court system's director of access to justice services, the team had to scale back its testing protocol. They settled on a refined list of 16 critical questions. This smaller set included queries AVA had previously answered incorrectly, complex scenarios, and common questions the team anticipated users would ask frequently.
This rigorous, human-led verification process is essential but also highlights a central paradox of automating such services. "It was just so very labor-intensive to do this," Marz admitted, noting the contrast with the popular narrative that generative AI will easily revolutionize public services.
The Future of AVA and Lessons Learned
Despite the setbacks, the Alaska Court System plans to launch AVA in late January. However, the project's goals have been adjusted based on the experience. The team no longer expects the chatbot to fully replicate the nuanced guidance a human facilitator can provide.
"We did shift our goals on this project a little bit," Marz said. "We’re not confident that the bots can work in that fashion, because of the issues with some inaccuracies and some incompleteness."
The ongoing nature of AI development also means the work is never truly finished. Martin explained that the team will need to conduct regular checks and potentially update the system as the underlying AI models evolve. It is not a "set it and forget it" solution.
The journey of AVA serves as a practical case study for public agencies nationwide. It demonstrates that while AI holds promise for increasing access to justice and improving efficiency, its successful and safe implementation requires a clear-eyed understanding of its current limitations, a significant investment in human oversight, and a cautious approach to its deployment.