Anthropic updated Claude’s “Constitution,” spelling out the chatbot’s “ethical character” and, crucially, requiring it not to “kill or disempower” humanity.
The Constitution is baked into Claude’s thinking; previous versions were softer guidelines, but as Claude becomes more powerful, its maker has added constraints on dangerous behavior.
Sci-fi fans might be reminded of Isaac Asimov’s Laws of Robotics, banning robots from harming humans, a near-impossible task with traditional programming because “harm” is ill-defined; natural-language AIs, though, can understand the instruction despite its fuzziness, making something like the Laws possible.
Of course, as in Asimov’s stories, there will still be tensions. Claude itself noted to Flagship that its instruction to be “honest” will sometimes conflict with its instruction to do no harm.