Align AI to Human Aspirations, Not Flaws, Argues New Paper

This paper presents a strong argument against the prevailing notion of aligning artificial intelligence with aggregated human preferences. The authors contend that human values, in their unfiltered pluralism, have historically led to societal issues ranging from inequality to political polarization, and therefore, AI should not inherit these potential flaws. Instead, the paper advocates for a foundational approach where AI is trained to adhere to a strict, non-negotiable set of objective alignment goals. These core principles include competence, factual accuracy, honesty, and lawfulness. The authors suggest that legitimate value tradeoffs and pluralism should only be incorporated at superficial levels, such as language style or contextual defaults, rather than at the fundamental level of values that could violate these objective standards. The paper highlights the dangers of unfiltered pluralistic values in AI and proposes four constructive commitments as an alternative. It also addresses several common objections to this approach, including commercial pressures, democratic legitimacy, and the potential for the "floor" itself to be culturally biased.