Anthropic has outlined measures and technical limitations governing use of its Claude language model ahead of U.S. federal, state and local elections on November 5, 2024. The company says 2024 is the first U.S. election cycle in which generative AI tools are widely available, and that since July 2023 it has taken concrete steps to detect and mitigate potential misuse and to direct users to authoritative election information.
Key rules and prohibitions
- Prohibiting campaigning and lobbying: Under Anthropic’s Usage Policy, Claude may not be used for political campaigning or lobbying. This includes promoting a specific candidate, party or issue, conducting targeted political campaigns, or soliciting votes or financial contributions.
- Combating misinformation and election interference: Use of Anthropic products to generate misinformation about election laws, candidates, or related topics is prohibited. The policy also forbids using Claude to target voting machines or obstruct vote counting or certification.
- Limiting outputs to text only: Claude does not generate images, audio, or video, which reduces the risk of election‑related deepfakes.
Detection and enforcement
Anthropic deploys a mix of technical and process controls to detect and prevent misuse:
- Automated systems enforced with human review to implement policy rules.
- Prompt modifications on claude.ai, auditing of use cases on Anthropic’s first‑party API, and in some extreme cases suspension of accounts.
- Collaboration with Amazon Web Services (AWS) and Google Cloud Platform (GCP) to detect and mitigate election‑related harms when users access Anthropic models on those platforms.
Testing, red‑teaming and continuous improvement
The company regularly conducts targeted red‑teaming to test how its systems respond to election‑related prompts. It also uses Policy Vulnerability Testing (PVT), a deeper testing process carried out with external subject‑matter experts, to identify risks. PVT focuses on misinformation, bias and adversarial misuse by posing relevant questions (for example, asking where and how someone can vote in the U.S.), documenting model responses, and noting whether safety interventions—such as declining to answer harmful queries—occur.
Anthropic has built automated evaluations to test systems at scale for a range of election‑related risks. These tests check, among other things:
- political parity in model responses across candidates and issues,
- the degree to which systems refuse to respond to harmful election queries,
- how robust the systems are at preventing misinformation and voter profiling tactics.
Findings from these assessments lead to policy updates, stronger enforcement processes, and technical refinements to the models themselves.
Providing accurate information and transparency
Because Claude’s training is not updated frequently enough to supply real‑time election details, Anthropic redirects users with election questions to authoritative, up‑to‑date voting information. The company implemented a pop‑up offering users the option to be redirected to TurboVote (a nonpartisan resource from Democracy Works) when they request voting information. TurboVote has recently been updated to include the names of all candidates running in federal and state elections, as well as ballot propositions.
Anthropic has also updated Claude’s system prompt to make the model’s knowledge cutoff date explicit so users understand the temporal limits of its training data.
Sharing learnings and industry engagement
Anthropic published some of the automated evaluations it developed and launched an initiative to fund third‑party evaluations that measure AI capabilities and risks. Throughout 2024 the company has engaged with global policymakers, civil society organizations and industry actors to discuss its election work and perform proactive scenario planning to prepare for potential election‑related abuse in the run‑up to election day.
Anthropic notes it cannot anticipate every possible misuse, but says it will continue to iterate on processes, testing and system improvements.
Related materials
- February 2024: Preparing for global elections in 2024
- May 2024: Updating our Usage Policy
- June 2024: Testing and mitigating elections‑related risks
- June 2024: Claude 3.5 Sonnet launch; Claude 3.5 Model Card Addendum
