Update README.md

README.md

@@ -115,7 +115,9 @@ The following hyperparameters were used during training:
 - warmup_ratio: 0.1
 - text_maxlength: 1024
 
-### Risk and Limitation
+### Bias, Risk and Limitation
 By replacing the retrieval component with a generative one, we lose the ability to update the knowledge state, which remains frozen to that of the LLM training time. This stands in contrast to the standard retrieve-then-read approach, which allows for the incorporation of new documents as new information becomes available. 
-Moreover, our method generates context for any given question, even when the medical LLM lacks knowledge. This may produce a noisy context with inaccurate or irrelevant information.
+Moreover, our method generates context for any given question, even when the medical LLM lacks knowledge. This may produce a noisy context with inaccurate or irrelevant information. 
+Unlike retrieve-then-read methods reliant on curated external corpora, using models to generate contextual documents may inadvertently amplify inherent biases and deviate from clinical and societal norms, potentially leading to the dissemination of convincing614
+medical misinformation.
 Therefore, we advocate for a cautious approach, recommending manual scrutiny of any output by domain experts before real-world utilization. This ethical precaution is vital to avoid disseminating potentially erroneous or misleading information, especially in the clinical and scientific communities