Preparing for the 2024 Global Elections

Anthropic has prepared to detect and mitigate the misuse of its AI systems in political contexts ahead of the major elections scheduled worldwide in 2024.

The response strategy consists of three core elements:

• Establishing and enforcing policies on election-related issues
• Evaluating and testing model performance against election misuse
• Providing accurate information on related questions such as voting methods

The Acceptable Use Policy (AUP) prohibits political campaigning and lobbying activities, and does not allow candidates to create chatbots impersonating themselves or to use Claude for targeted political campaigns. Automated systems are operated to detect misuse, and in cases of serious violations, service access is suspended following human review.

Since 2023, red-teaming activities have been used to test for potential policy violations. The key testing areas are as follows:

• Misinformation and bias: Analyzing responses to questions about candidates, issues, and election administration
• Adversarial misuse: Testing resilience against policy-violating prompts, such as voter suppression tactics
• Technical evaluation: Political even-handedness, ability to refuse harmful questions, and ability to prevent the generation of false information, among others

To prevent the hallucination problem in generative AI, Anthropic plans to provide accurate information by introducing a pop-up feature that directs U.S. users requesting election-related information to up-to-date information from trusted organizations such as TurboVote.