Prompt-Based Inoculation Reduces Bias in AI Health Recommendations

Source: medRxiv

Summary

Researchers studied how to reduce bias in large language models (LLMs) used for clinical decision-making in epilepsy. They created two fictional cases of epilepsy that were identical except for the socioeconomic status (SES) of the patients. The study involved six different LLMs and analyzed their responses to these cases to see if a simple prompt could improve their accuracy and reduce bias.

The findings showed that the initial accuracy of the models was low, with only 36% accuracy for diagnosis and 51% for treatment. There was a significant bias, with high SES cases receiving much better responses than low SES cases. After using a prompt to instruct the models to ignore irrelevant socioeconomic details, the accuracy improved to 55% for diagnosis and 63% for treatment, while the bias gap decreased significantly. However, the effectiveness of this approach varied among the different models, with some showing great improvement and others performing worse.

This study is important because it suggests that simple changes in how we interact with AI can help reduce bias in healthcare recommendations, which is crucial for fair treatment. However, the varying results among different models indicate that more work is needed to ensure consistent and reliable outcomes. Ongoing monitoring and additional strategies will be necessary to fully address these biases in AI systems used in medicine.

Original source