Researchers identify biases and empathy patterns in large language models

New research highlights how training data influences the behavior and accuracy of large language models (LLMs). A study led by Georgia Tech researchers found that the tone of an AI's output is closely linked to its training sources. Models that adopt an empathetic tone tend to rely more on interpersonal dialogue, such as FAQs and customer service transcripts, whereas models with an objective tone prioritize neutral sources like textbooks.

Separately, a collaborative study by Osaka Metropolitan University and MD Anderson Cancer Center revealed that LLMs exhibit significant bias in medical recommendations. When asked to suggest FDA-approved drugs, the models overwhelmingly favored brand-name products that are highly visible in search engines or have dedicated Wikipedia pages.

The study noted that these models often failed to recommend more recent FDA approvals or treatments commonly used in actual physician practice. If physicians were to follow these AI-driven recommendations, researchers estimate annual acquisition costs in the US could increase by approximately $55.6 million, representing a massive spike in patient costs.