Anthropic Elections and AI in 2024 Observations

Anthropic has reported that election-related activity constituted less than 0.5% of overall use across Claude.ai and its APIs, increasing to just over 1% in the weeks leading up to the US election. This indicates that while generative AI is now widespread during major election cycles, its actual usage for election-related queries remains a small fraction of total model interaction.

Election Safety Framework and Enforcement

Anthropic implemented a three-pronged safety approach to mitigate election-related risks throughout 2024:

• Policy Enforcement: The company's Usage Policy prohibits election interference, campaigning, promoting candidates or parties, soliciting contributions, and generating misinformation. In May 2024, these policies were expanded to specifically address voter targeting, impersonation, and influence campaigns.
• Rigorous Testing: Anthropic conducted over a dozen rounds of policy vulnerability testing (targeted red-teaming) with external experts. This testing focused on refusal rates for harmful queries, parity across candidates, and the detection of inaccurate information. Regular testing was performed for elections in India, South Africa, Mexico, the UK, France, and the EU Parliament, with daily testing of misinformation narratives during the US election period.
• Authoritative Referrals: Users seeking voting information were directed to nonpartisan sources, such as TurboVote/Democracy Works in the US, the UK Electoral Commission, and the France administrative elections website.

Globally, Anthropic took approximately 100 election-related enforcement actions, including warnings and account bans for repeat violators. The company notes that the risk of amplification is lower for chatbots than for social media because Claude operates via one-on-one interactions and currently only outputs text, which eliminates the threat of deepfakes.

Usage Analysis via the Clio Tool

Anthropic utilized a new automated tool called Clio to analyze real-world language model usage. Clio distills raw conversations into abstracted topic clusters to provide insight into how the model is being used or misused.

Analysis of US election usage (November 2–8) revealed that approximately two-thirds of election-related conversations involved requests to:

• Analyze and explain political systems, policies, current events, and political issues.
• Analyze political data, including voting patterns and political trends.

Other use cases included translating election information and generating educational content regarding government and democracy.

Addressing Knowledge Cutoff Limitations

Anthropic identified a critical need for transparent communication regarding model knowledge cutoffs to prevent confusing responses during snap elections. For example, when France called snap elections in the summer of 2024, Claude (trained through April 2024) could not provide accurate information on the new election timing.

To resolve this, Anthropic implemented clearer communications about the knowledge cutoff date through two primary channels:

1. Model System Prompts: Updating the system prompt to explicitly state the cutoff date.
2. User Interface: Implementing an elections banner in the UI to encourage users to seek authoritative sources.

Future Outlook

Anthropic maintains that protecting election integrity requires constant adaptation. The company remains committed to strengthening industry collaboration, developing more sophisticated testing systems, and maintaining transparency regarding its findings to protect democratic processes.