Anthropic Collaborates with Public to Draft AI Constitution
Anthropic, an AI startup known for its Claude chatbot, has partnered with the Collective Intelligence Project (CIP) to develop a constitution for an AI system. To create this constitution, Anthropic sought input from approximately 1,000 American citizens. The company currently uses its own curated constitution for its chatbot, which is based on Constitutional AI (CAI), a method developed by Anthropic to ensure that large language models (LLMs) adhere to high-level normative principles. The new constitution draws inspiration from documents like the Universal Declaration of Human Rights published by the United Nations.
Publicly Sourced Constitution and Experiment Results
In a recent blog post, Anthropic revealed the publicly sourced constitution resulting from the consultation process and shared details about training a new AI system using the CAI method. The company explained that the experiment aimed to explore how democratic processes can influence AI development. During the consultation, Anthropic and CIP used Polis, a platform for gathering public opinions, to ask participants to vote on existing principles or suggest their own.
While there was roughly a 50% overlap between the publicly sourced constitution and Anthropic’s in-house version, some public principles did not closely align with the company’s original document. For example, one principle suggested prioritizing balanced and objective information that reflects all sides of a situation, while another emphasized being understanding and adaptable towards people with disabilities.
Conflicting statements among opinion groups also arose during the consultation. One group believed that AI should prioritize the collective or common good over individual preferences or rights, while another group argued for prioritizing personal responsibility and individual liberty over collective welfare.
The Impact of Public Input on AI Behavior
CIP’s evaluation of the experiment concluded that the public model displayed less bias in relation to various stereotypes and performed equally well to the baseline model in areas such as math, natural language understanding, and helpfulness. The organization emphasized that if generative AI is going to shape how people work, communicate, and interact on a large scale, it is crucial to have public input into the behavior of these models.
Hot Take: Harnessing Public Input for AI Development
Anthropic’s collaboration with the public to draft a constitution for an AI system showcases the potential of democratic processes in shaping technology development. By involving around 1,000 American citizens in the consultation process, Anthropic and CIP aimed to ensure that AI systems adhere to high-level normative principles. While there were differences between the publicly sourced constitution and Anthropic’s original document, this experiment highlights the importance of public input in reducing bias and ensuring AI models perform effectively. Moving forward, incorporating democratic principles into AI development could lead to more ethical and inclusive technology.
By
By
By
By
By