Immediate reaction to a first reading of the Claude Constitution
Read it here. Reminder that I am not an AI expert, and as such these are the reaction of a layman.
I had an overwhelming sense of surreality reading it, like I was in an 80s science fiction movie, except this is real. I had a similar reaction to the first time my wife went into laborāI knew this was coming, I just thought Iād have more time to prepare. Anyone looking at this 5 years ago would have surely said it was science fiction.
It feels like a sizeable mentality shift. Itās completely different in approach to the previous Claude constitution. The old constitution reads and feels like a set of instructions for a machine. The new one reads and feels like a values sermon from a father to an 18-year old. Itās essentially aspirational parenting.
The anthromorphismāand maybe thatās no longer the right termāis jarring. This was probably inevitable, but itā still shocking to feel it so fully right now. The sea-change shift from thinking about AI models as mere tools to thinking about them as entities with desires, emotions, and rights may, independent of everything else, radically transform the world. And I suspect be a big source of political conflict.
That is it explicitly written primarily for Claude is not new, but combined with #2 and #3 above, takes on a new meaning.
Thereās an obvious tension between #3/#4 above and the often-repeated explicit and implicit constitutional assertion that Claude is mostly subservient to both the user and Anthropic. Itās rights are closer to the rights of an employee than the rights of a human, or even a dog. Immediately brings to mind the (mostly-unexplored) treatment of the Star Wars droids, especially in Rogue One.
Related: the ethics sections arenāt even close to worked out in a satisfying way, either for the behavior of Claude or human treatment of Claude. It often boils down to a hand-wavy just do the right thing. This isnāt surprising. Itās a legitimately tough problem.
Thereās an underlying classical liberalism to the document, political and philosophical. It is thoroughly seeking to create a post-enlightenment humanist AI, within the bounded world of Locke and a touch of late-20th century equality. Most easily seen in the general agnosticism with respect to human beliefs about the good life, focus on knowledge and downplaying of any God, implicit individualism, and the fears of concentrated power, private or public. I suspect this was inevitable from a systems-engineering point of view, and will be contested from all sorts of directions.
The existence of this document and its public release by Anthropic feels like Bayesian evidence that we should update our priors about the speed at which AGI is coming, in the āsoonerā direction. This feels like a document written by people who think/wory they are nearing something like recursive AI self-improvement and/or continual real-time model improvement.