The Nevada state employment department has revealed it will be utilizing Artificial Intelligence to help speed up its unemployment appeals process by analyzing transcripts from appeal hearings and issuing a recommendation decision.
The department reported it has been ‘buried’ in a backlog of claims since the start of the pandemic, and is desperately looking for ways to get back on track.
It has also confirmed it will not be training a new generative AI model for the system, but will instead use Google’s Vertex AI studio, which will reportedly cut the review process from hours down to just five minutes – in spite of new research which suggests AI models in general are worse than humans in ‘every way’ at summarizing documents, and often create additional work for workers.
Clearing the backlog
Experts have warned against the approach, not just because Large Language Models don’t understand text or reason logically and within a context the way that humans can, but also because it may not save the department much time.
“If someone is reviewing something thoroughly and properly, they’re really not saving that much time,” noted Morgan Shah, Director of Community engagement for Nevada Legal Services. “At what point are you creating an environment where people are sort of being encouraged to take a shortcut?”
A lack of accuracy in the model concerns Nevada Legal Services lawyers, who cite AI ‘hallucinations’ as a worry, which is an industry term used to describe when an AI model produces factually incorrect or misleading responses.
Any AI decision will also be double checked and reviewed by a human referee before it is handed out – but if the human referee makes a decision based on the AI hallucination, a court may not be able to overturn the decision.
The infamous IBM quote springs to mind, ‘a computer can never be held accountable, therefore must never make a management decision’. Research has shown many of us are still very wary about AI, especially in high-risk products (like medical diagnoses and automated vehicles). The success of this experiment could have an impact on a wide range of government departments going forward.
Via Gizmodo
