AI
New benchmark measures AI safety risks in high-stakes professions
Researchers introduce Trident-Bench to evaluate how well language models handle ethical challenges in law, finance, and medicine.
Large language models are increasingly making their way into sensitive professional fields—law, finance, and medicine—where mistakes carry serious consequences. Yet most safety testing for these systems has focused on general harms, missing domain-specific risks that matter in regulated industries. A new research effort aims to change that.
Building domain-specific safety standards#
Researchers have developed what they call Trident-Bench, a evaluation framework rooted in actual professional ethics codes. The benchmark draws its principles directly from established industry standards: the American Medical Association's framework for medical ethics, the American Bar Association's professional conduct rules, and the CFA Institute's code of conduct for finance professionals. This grounding in real-world regulatory expectations ensures the benchmark tests what actually matters in each field.
Testing 19 AI models reveals mixed results#
The team evaluated 19 different language models—ranging from widely-used generalist systems like GPT and Gemini to domain-specific variants. The findings paint a nuanced picture. Larger general-purpose models often clear basic safety thresholds, but domain-specialized models frequently stumble on the subtler ethical judgments that professionals routinely navigate. This gap suggests that simply fine-tuning a model for a particular industry isn't enough to ensure it makes sound ethical decisions.
Why this matters for professional AI deployment#
As organizations consider deploying AI systems to handle real work in law, medicine, and finance, the stakes couldn't be higher. An AI-drafted contract with hidden risks, a diagnostic system that overlooks alternatives, or financial advice that ignores conflict-of-interest rules could all cause genuine harm. Trident-Bench provides one of the first systematic ways to measure these specific vulnerabilities before deployment, laying groundwork for safer AI integration in professionally regulated fields.
