A well-intentioned project to help grieving Alaskans navigate probate paperwork has turned into a cautionary tale about the challenges of deploying AI in government services. The Alaska Court System launched development of the Alaska Virtual Assistant, or AVA, over a year ago with ambitious goals: create an AI chatbot that could guide residents through the complex maze of forms and procedures required to settle a deceased person’s estate.
What officials envisioned as a three-month project has stretched into a grueling 15-month ordeal marked by technical setbacks and troubling inaccuracies.
“With a project like this, we need to be 100% accurate, and that’s really difficult with this technology,” said Stacey Marz, administrative director of the Alaska Court System. The stakes are simply too high for the usual approach of launching a minimum viable product and fixing problems later, she explained.
The AVA experience reveals why government agencies nationwide are moving cautiously with AI despite the technology’s hype.
The Law School that Doesn’t Exist, The Battle of Alaska with AI Hallucinations
A recent Deloitte report found that less than 6% of local government practitioners prioritize AI as a service delivery tool. The gap between soaring AI investment and limited real-world adoption becomes clearer when you examine what Alaska’s team has encountered.
One of the most persistent problems has been hallucinations, instances where the AI confidently provides false information. Aubrie Souza, a consultant with the National Center for State Courts who worked on AVA, recalled a particularly troubling example. When users asked where to find legal help, the chatbot suggested contacting alumni from Alaska’s law school. The problem? Alaska doesn’t have a law school.
“We had trouble with hallucinations, regardless of the model, where the chatbot was not supposed to actually use anything outside of its knowledge base,” Souza said. The team has worked extensively to constrain AVA to only reference approved probate documents rather than conducting wider searches that could introduce errors.
The consequences of such mistakes go far beyond embarrassment. People relying on incorrect information about probate could make costly errors affecting inheritances, property transfers, and family estates. Marz noted that inaccurate or incomplete guidance “could be incredibly damaging to that person, family or estate.”
Why AVA Swapped Condolences for Rule-Following?
Building AVA has required countless careful decisions about everything from the AI’s personality to its tone with grieving users. Early versions were programmed to express empathy, but user testing revealed this backfired. People dealing with death didn’t want yet another expression of sympathy, especially from a machine.
“Through our user testing, everyone said, ‘I’m tired of everybody in my life telling me that they’re sorry for my loss,'” Souza explained. The team removed those condolences entirely.
Tom Martin, a lawyer and law professor who developed AVA’s technical systems through his company LawDroid, said choosing the right AI model involves considering their distinct “personalities.” Some models focus on rule-following while others try to show off their knowledge. For legal applications, rule-following is essential.
To evaluate AVA’s reliability, the team initially created 91 test questions covering various probate scenarios. But reviewing answers proved too time-consuming given the need for careful human verification. They eventually narrowed the test to 16 critical questions that balanced complexity with practicality.
Even after launch, the work won’t stop. Because AVA runs on systems like OpenAI’s GPT models that constantly evolve, the court system must commit to ongoing monitoring. Martin acknowledged this reality: “We’ll need to stay on top of it rather than a purely hands-off situation.”
Cost remains a silver lining. Martin noted that 20 AVA queries cost only about 11 cents under their current setup, far cheaper than staffing a help line.
Alaska’s Long Road from AI Hype to Helpful Tool?
Despite the struggles, AVA is scheduled to launch in late January. But Marz’s goals have shifted. Rather than fully replicating human facilitators at the self-help center, she now sees AVA as a useful supplement that may improve as AI technology advances.
“It was just so very labor-intensive to do this,” Marz reflected, “despite all the buzz about generative AI, and everybody saying this is going to revolutionize self-help and democratize access to the courts. It’s quite a big challenge to actually pull that off.”
Alaska’s experience offers a reality check for governments eyeing AI solutions. When accuracy matters and real harm is possible, the path from AI hype to helpful tool is longer and harder than the headlines suggest.
