When I asked it to double check, it immediately apologised and corrected its error. This was a tiny detail that reflected a larger issue. In a world where the public is already well aware of traditional media echo chambers and social networking algorithmic bias, AI output is often associated with neutrality and objectivity.  A large language model output is, in fact, a data driven response to whatever query you offer but the context surrounding the text is more complex than it appears. A data driven result from a large language model is only as objective as the data on which it is trained. Large language models learn from enormous datasets scraped from both current internet sources as well as archives of historical texts.  The curation of these datasets prioritizes sources with a high volume of digitized material which reflects in its over representation of western language and culture. The archives used as training data contain systemic biases that reflect humanity's most deep seeded prejudices like sexism and racism. The chatbots therefore reflect and amplify these patterns weaving biases of the past into the responses of the future.

The mere awareness of systemic bias underpinning the veneer of AI’s scientific accuracy is important, but it is only half the battle. To counteract the inherent prejudice from tainting each output, there are efforts to create guardrails like policies defaulting to gender neutral language in certain contexts as well as active initiatives seeking to counteract the imbalances throughout the training process and within the dataset itself. One of these efforts is championed by the MIT Media Lab and is called Perspective-Aware AI.  PAi is a framework that actively models, codes, and balances a rich variety of human viewpoints in an effort to make AI more inclusive, transparent, and fair. The system builds digital profiles which are referred to as "chronicles” that highlight overlooked groups to mathematically force the AI to take historically underrepresented perspectives into account. As AI continues to engrain itself into our world and inform our perspectives, the herculean effort to reveal and counteract AI biases based on historical training data is a necessary task in order to avoid projecting the prejudices of the past onto the social fabric of the future.