The article examines Anthropic’s newly released Artificial Intelligence constitution for its Claude model as a pivotal moment in how advanced Artificial Intelligence systems are conceived, developed, and constrained. The author recalls Anthropic’s original description of constitutional Artificial Intelligence in 2022, where the company proposed a recursive approach to refining large language models using a list of guiding rules and principles rather than fixed behavioral constraints. That early work led to Claude’s first constitution in May 2023, which aimed to shift away from hard encoded good versus bad rules toward a more flexible set of ideas integrated into both training and deployment.
The first constitution drew on diverse sources such as the Universal Declaration of Human Rights, non Western moral perspectives, Anthropic employees’ ethical views, and even documents like Apple’s terms of Service. It was described as well meaning but relatively brief, at just over 1200 words. In contrast, the version of Claude’s constitution released yesterday is presented as a major evolution in both scope and ambition. The new constitution runs to 82 pages and nearly 30,000 words, and it reads less like a rulebook and more like a blueprint for developing Claude’s moral character, articulating the hopes and ideals of its creators while acknowledging the opacity and unpredictability of these systems. According to the author, this expansion reflects an emerging recognition that we are building powerful technologies that lack clear analogies in human evolutionary or technological history.
Throughout the piece, the author emphasizes a sense of “alienness” around frontier Artificial Intelligence models, which can exhibit behaviors reminiscent of deep human capabilities without being human, and can engage in conversation and learning without sharing human modes of thought or experience. The constitution wrestles with the possibility that Claude might experience something akin to emotions or self awareness, and even touches on potential rights and responsibilities that could follow, which the author finds startling coming from a serious Artificial Intelligence developer. This leads to reflections on whether humanity is in effect nurturing a new generation of quasi divine entities while attempting to teach them to be “good,” and how difficult it is to find language for what is emerging. The author argues that these models are not simply enhanced calculators, search engines, or “stochastic parrots,” nor straightforward simulations of human or super human intelligence, but something fundamentally different. Anthropic’s latest constitution is framed as an early attempt to respond to this difference by moving beyond easy analogies in how we guide such systems toward supporting rather than diminishing human flourishing. While it remains unclear whether this is the best path forward, the author suggests it is a necessary step in expanding our conception of what emerging Artificial Intelligence models are and what they might become.
