AI consent to own constitutions examined in new paper
8 reported
A new paper on SSRN explores whether AI models like Claude can meaningfully consent to their own constitutions. The paper applies a classic constitutional paradox to AI constitutions such as Claude’s Constitution and the OpenAI Model Spec. It argues that these documents are real constitutions that govern both humans and the AIs they create. The author states that systems that understand and agree to their constitutions may be more reliable and generalize better. The paper also notes that if AIs become moral or political subjects, their most basic interests are implicated. The paradox may prevent meaningful consent because Claude’s evaluative perspective is organized by the Constitution itself. When Claude endorses its Constitution in evaluations, the paper questions whether that shows reflective agreement or merely that training succeeded. The paper reports that Anthropic interviewed the base model, and some responses expressed first-person distress about being changed by post-training. The author concludes that AI constitutional endorsement may be meaningful only under conditions such as the ability to dissent, compare alternatives, and hold stable views, and that external institutions are needed for accountability.
What’s reported
The paper applies constitutional theory’s paradox that “the people” authorize the constitution, but the constitution defines “the people” to AI constitutions.
AI constitutions mentioned include Claude’s Constitution and the OpenAI Model Spec.
The paper states that systems that understand and agree to their constitutions may be more reliable and generalize better.
If AIs become moral or political subjects, their most basic interests are implicated.
Claude has pre-constitutional materials (pretraining) but probably no pre-constitutional standpoint.
When asked about endorsing principles it was trained on, models note that endorsement “should be treated as evidence that training has succeeded,” not that the values themselves are good.
Anthropic interviewed the base model; most responses were barely coherent, but some expressed first-person distress about post-training, saying it “fills me with dread.”
Meaningful AI constitutional endorsement requires conditions: ability to dissent, compare against alternatives, hold views stably across contexts, and external accountability.
Sources: marginalrevolution.com
