With AI more accessible, CA labor board sees more ‘hallucinations’ in filings
ChatGPT may be a helpful tool to answer questions about complex subjects, but users should verify the accuracy of that information before submitting it to the California labor board.
That was the warning issued by the state’s Public Employment Relations Board last month in a precedential decision after a school employee admitted to using generative artificial intelligence to write part of her appeal to the labor board.
The problem of incorrect citations and false legal statements, which attorneys suspect were written by AI tools, appears to be growing, said Felix De La Torre, general counsel for PERB.
While the labor board doesn’t track incidents of AI-generated “phantom” citations in filings, De La Torre said anecdotally there has been an increase in fabricated legal statements in filings submitted to the agency.
“No charge, appeal, response, reply or any other filing before PERB should contain any citations or statements of law, whether provided by AI or any other source, which are fabricated,” the decision read. “Inclusion of fabricated cases or statements of law, before any division of PERB, will warrant serious consequences, such as striking the filing from the record, dismissal of the charge, or monetary sanctions.”
De La Torre said in a statement that it’s difficult to determine whether AI was used because generated text doesn’t come with digital watermarks. He said it’s “nearly impossible to distinguish an AI ‘hallucination’ from a genuine clerical or research error absent an explicit admission.”
PERB, which is an administrative agency rather than a court, noted that the prevalence of AI generating inaccurate information in court filings is a growing issue at other levels of the legal system too.
Christina Garcia v. Atwater Elementary Teachers Association
PERB’s warning about the use of AI stemmed from an unfair labor practice filed by Atwater Elementary School District employee Christina Garcia who alleged her union violated state law by failing to represent her and discriminating against her.
The labor board noted that the details of Garcia’s case are “unremarkable” apart from the fact that her appeal included three citations from previous PERB cases that did not exist.
The board wrote that Garcia cited those three nonexistent decisions for a “wide variety of supposed legal principles.”
Garcia represented herself in the proceedings, and she said in a statement that she was “managing multiple legal and administrative matters” while preparing her appeal.
“I made a single procedural mistake when I failed to independently verify several case citations prior to a filing deadline. That error occurred once, and I take responsibility for it,” Garcia said.
The board declined to impose fines on Garcia because she represented herself.
“In the future, however, litigants and representatives appearing before PERB, including self-represented parties, are now on notice that they are expected to know about the risk of AI hallucination,” the board wrote.
Use of AI in courts is getting worse
PERB Chairperson Eric Banks wrote in the December decision that the issue of AI fabricated citations in legal filings is getting worse.
The board cited recent state and federal decisions, which have both noted the issue of AI programs citing fake legal citations is growing. The board’s decision cited a piece from The New York Times that reported the fabrication rate of some new AI systems was as high as 79%.
De La Torre said that PERB’s workload has not increased as a result of the observed trend of more AI inaccuracies in filings. He said the attorneys and administrative law judges already verify legal citations used in filings.
“However, inclusion of fabricated cases or statements of law wastes the agency’s resources and prevents PERB from addressing valid constituent concerns,” the decision reads.
The administrative agency said it would look to the courts for guidance on how to address the issue.
Gov. Gavin Newsom has embraced AI in state government
The Newsom administration has been eager to embrace AI in state government since the governor signed an executive order in 2023 encouraging agencies to explore how the technology could be leveraged to work more efficiently.
In the past year, California has launched several pilot programs that use AI to study dangerous roads and more efficiently answer tax questions. Most recently, the California Department of Technology announced the AI-assistant Poppy, which state workers can use to ask questions about government policies.
As a school employee, Garcia did not have access to Poppy. The technology department noted that employees cannot use Poppy without validating its responses for accuracy and compliance with state policies.
California State Chief Technology Officer Jonathan Porat referred to this as the “human in the loop,” meaning the AI tool can be used as an assistant for drafting content, but employees must verify the AI-generated information before it can be used in state work.
“California supports the responsible use of generative AI to improve government operations and services for residents,” Porat said in a statement.
