Anthropic details safeguards for Claude ahead of U.S. elections

Anthropic outlined policies, technical controls and testing intended to limit election-related misuse of Claude ahead of federal, state and local elections in the United States on November 5, 2024.

The company said it had taken steps since July 2023 to detect and mitigate potential misuse of its tools and direct users to authoritative election information. Its announcement summarized that work, including policy restrictions, enforcement practices, evaluations and voting-information redirects.

Restrictions on political uses

Anthropic said a May update to its Usage Policy clarified prohibited election and voting uses. The company prohibits using its products for political campaigning and lobbying. Under that policy, Claude cannot be used to promote a candidate, party or issue, conduct targeted political campaigns, solicit votes or seek financial contributions.

The policy also prohibits generating misinformation about election laws, candidates and related subjects. Anthropic said Claude may not be used to target voting machines or obstruct the counting or certification of votes.

Claude produces text rather than images, audio or video. Anthropic said this limitation eliminates the risk of election-related deepfakes generated through the product.

Enforcement and testing

Anthropic said it deploys automated systems to enforce its policies and audits those systems through human review. Its methods include modifying prompts on claude.ai, auditing use cases on its first-party API and, in extreme cases, suspending accounts.

The company also works with Amazon Web Services and Google Cloud Platform to identify and mitigate election-related harms involving users who access Anthropic models through those platforms.

Anthropic said it regularly conducts targeted red-teaming focused on election prompts. It also uses Policy Vulnerability Testing, conducted with external subject matter experts, to examine misinformation, bias and adversarial abuse. That process includes selecting relevant questions, documenting model responses and recording whether safety interventions occur, such as refusing harmful requests.

Automated evaluations test the systems at scale, according to Anthropic. The tests examine political parity across candidates and topics, refusals of harmful election queries, and resistance to misinformation and voter-profiling tactics. Anthropic said findings from this work inform changes to its policies, enforcement processes and models.

Voting information and transparency

Anthropic said its models are not trained frequently enough to supply real-time election information. For election-related questions, the company redirects users toward current, authoritative voting resources.

A pop-up on Claude gives users asking for voting information the option to visit TurboVote, a nonpartisan resource from Democracy Works. Anthropic said TurboVote had recently added the names of candidates running in federal and state elections, along with ballot propositions.

The company also updated Claude’s system prompt to state its knowledge cutoff date. Anthropic said it released some automated evaluations used in its election work and started an initiative to fund third-party evaluations of AI capabilities and risks.

Anthropic said it met during the year with policymakers, civil society organizations and industry participants, and conducted scenario planning for possible election-related abuse before U.S. election day.

Source: Anthropic’s “U.S. elections readiness” announcement, published October 8, 2024.