Suliman: Protecting Mass. from AI hallucinations

Massachusetts is taking steps to address the risks of AI hallucinations after recent court cases revealed lawyers using AI-generated filings that cited nonexistent legal precedents. Judges have sanctioned attorneys for relying on inaccurate AI outputs, highlighting the need for accountability. The issue extends beyond legal filings, as AI systems have fabricated medical content, research papers, and police reports, often producing errors that experts can detect but average users cannot. Governor Maura Healey recently announced a partnership with Google to provide free AI training for state residents and launched a ChatGPT-powered AI assistant for state agencies, making Massachusetts the first in the U.S. to deploy such technology across executive branches. However, the state’s policy acknowledges that generative AI systems can produce unreliable data, emphasizing the need for consistent validation to ensure accuracy. Legislation proposed in Massachusetts aims to require AI creators to disclose uncertainty in responses and cite sources used for generation. The state also advocates for mandatory warnings about AI hallucinations and independent third-party oversight to assess inaccuracies. Users should be able to report errors, with AI companies obligated to publish periodic reports on how feedback improves models. Critics argue that AI hallucinations—where systems fabricate information—may be inherent to their design, complicating efforts to eliminate the problem entirely. Some AI companies already include disclaimers about potential inaccuracies, but enforcement remains inconsistent. Massachusetts’ approach focuses on mitigating harm through legal measures rather than waiting for technological solutions. The urgency was underscored by a recent case where an immigrant relied on a free AI chatbot for legal advice before consulting a professional after being warned of potential inaccuracies. The state’s initiatives reflect broader concerns about AI adoption without safeguards against misleading outputs.