Writing an AI Constitution
A workshop exploring the values and principles of AI through collaborative drafting.
Shown here installed at Tate Modern, developed as part of a residency funded by Anthropic. Delivered in collaboration with Nikos Kourous.
Anthropic pioneer a method of training LLMs named Constitutional AI. Designed to replace Reinforcement Learning from Human Feedback, Constitutional AI employs an extra LLM trained on a pre-determined set of values and principles to censor the initial model's outputs, reducing the need for human oversight.
But who gets to write these values? At the moment, a handful of people in Silicon Valley.
During the workshop, visitors were invited to discuss and reflect on the Constitution currently being used by Anthropic, and encouraged to alter it, amend it, and add new clauses – all with the aim of drafting a more equitable and open set of values for how AI should behave going forward.
