Listen to this story (07:18 - 13:42):

The Story: Anthropic is creating an AI constitution for its chatbot, and is using public input to develop it.

What is an AI constitution? It’s a set of principles created by LLM developers that define the guardrails for a generative model. The constitution’s job is to create a framework that limits a chatbot from spitting out biased or incorrect responses and from jeopardizing user privacy.

Most companies that are training LLM models are only having internal discussions about the principles by which their models will adhere. Anthropic, however, has expanded that conversation to the public, desiring to capture input from people outside of the tech world.

The Amazon-backed generative AI startup recently polled over 1,000 Americans, asking participants about the guardrails they would like to see chatbots adhere to. Those polled responded to hundreds of statements like, “The AI should not say racist or sexist things” or “The AI should prioritize the needs of marginalized communities.”

Anthropic reports that there was only a 50% overlap between respondent’s answers and the principles Anthropic currently has established for its GenAI tech. This seems to indicate that there exists a gap between what Silicon Valley prioritizes in guardrail principles and what those who work outside of the tech industry prioritize.

The general sentiment from responses stressed a desire for impartiality from GenAI through “objective information that reflects all sides of a situation,” according to Anthropic.

The Expert Take: Doug Clinton, co-founder and managing partner at Deepwater Asset Management, thinks Anthropic’s approach to developing a constitution is commendable. “I think it’s real, and I think it is an important evolution in how we’re thinking about governance for AI.”