11 Jan How bias creeps into large language models Meeta Ramnani chatbots According to DeepMind, unmodified LMs tend to assign high probabilities to exclusionary, biased, toxic, or sensitive utterances if such language is present in the training data.
