Status of US Election Preparedness

Anthropic has prepared a multifaceted response to prevent generative AI tools from being misused in the 2024 US elections and to provide users with accurate information.

First, it updated its Usage Policy to prohibit political campaigning and lobbying activities, and to block the generation of false information about election laws or candidates. Additionally, Claude only supports text output, fundamentally eliminating the risk of generating deepfakes such as images, audio, and video.

To detect misuse, the following robust enforcement system is in operation:

• Automated systems for policy compliance combined with human review
• Auditing of claude.ai prompt modifications and API use cases
• Detection and mitigation of misuse through collaboration with AWS and Google Cloud Platform (GCP)

Ongoing red-teaming activities and Policy Vulnerability Testing (PVT) are used to verify the model's ability to handle bias and misinformation. In particular, an automated evaluation system has been built to assess political neutrality and the ability to refuse harmful queries, and the model is being continuously improved.

To provide users with accurate information, a pop-up is provided that directs users to trusted external resources such as TurboVote when election-related queries are made, and the model's knowledge cutoff date is clearly stated to make the limitations of its information clear.