Overview / Section 6

Ambiguous Implicit Cultural Bias

Dario proposes, “AI could be used to both aggregate opinions and drive consensus among citizens, resolving conflict, finding common ground, and seeking compromise.” Much of the pragmatic appeal of LLMs is that given enough training data, they generalize solutions very well.

Unfortunately, the preponderance of original human-generated text data fed into frontier models comes from white, affluent men. This is due solely to the cultural advantages that white, affluent men have possessed over the course of recorded history, that created the opportunity for them to dominate white-collar jobs, media, positions of power, and therefore generate of lot of text.

U.S. Federal law considers the following protected classes: race, color, religion, sex, national origin, age, disability, genetic information, and pregnancy. By law, none of that metadata on a person can be used to discriminate against them in a protected setting. For AI, the weight that training data gives to white, affluent men creates an implicit bias in the generalization of output produced by the LLMs. This bias reinforces output and ultimately solutions that run counter to those class protections under the law.

If Anthropic’s mission “remain beneficial to humanity” applies to all of humanity, then this implicit bias must be studied, qualified, and if possible, somehow countered or removed. It seems like a constitution would be a good place to make this statement and at least address the goal of countering that implicit bias. Otherwise, it is difficult to imagine Claude can generalize well enough to help us transition to a truly egalitarian future.