Election Safeguards Update

Ahead of major elections worldwide, including the US midterm elections, Anthropic has strengthened safeguards to ensure Claude can provide political information accurately and neutrally.

To prevent political bias, the model has been trained according to Claude's Constitution principles to address diverse political perspectives with equal depth and analytical rigor. This is implemented through character training and system prompts, and fairness across the political spectrum is evaluated before each model release. Evaluation results showed Opus 4.7 scoring 95% and Sonnet 4.6 scoring 96%.

In addition, the following defense systems are in place to prevent election-related misuse:

• Usage Policy: Strictly prohibits deceptive campaigns, generation of fake content, voter fraud, and the spread of false information related to voting.
• Detection and Enforcement: Automated Classifiers and a dedicated threat intelligence team detect and block manipulated election interference attempts in real time.

In a recent test of 600 prompts, Opus 4.7 demonstrated a 100% appropriate response rate in complying with election-related policies and refusing harmful requests, while Sonnet 4.6 recorded 99.8%, proving strong defense capabilities.