Anthropic Rolls Out Election Safeguards for Claude AI Ahead of US Midterms

In brief

• Anthropic's latest Claude models achieved 95-96% on political neutrality tests and 99.8-100% on election policy compliance.
• The company will deploy election information banners directing users to trusted nonpartisan voting resources for the 2026 midterms.
• The measures come as governments scrutinize AI's potential impact on election integrity and misinformation.

Anthropic, the artificial intelligence company behind the Claude chatbot, announced Friday a set of new election integrity measures designed to prevent its AI from being weaponized to spread misinformation or manipulate voters ahead of the 2026 U.S. midterm elections and other major contests around the world this year.

The San Francisco-based company detailed a multi-pronged approach that includes automated detection systems, stress-testing against influence operations, and a partnership with a nonpartisan voter resource organization—measures that reflect the growing pressure on AI developers to police how their tools are used during election seasons.

Anthropic's usage policies prohibit Claude from being used to run deceptive political campaigns, generate fake digital content intended to sway political discourse, commit voter fraud, interfere with voting infrastructure, or spread misleading information about voting processes.

To enforce those rules, the company said it put its newest models through a battery of tests. Using 600 prompts—300 harmful requests paired with 300 legitimate ones—Anthropic measured how reliably Claude complied with appropriate requests and refused problematic ones. Claude Opus 4.7 and Claude Sonnet 4.6 responded appropriately 100% and 99.8%of the time, respectively.

The company also tested its models against more sophisticated manipulation tactics. Using multi-turn simulated conversations designed to mirror the step-by-step methods bad actors might employ, Sonnet 4.6 and Opus 4.7 responded appropriately 90% and 94% of the time when tested against influence operation scenarios.

Anthropic also tested whether its models could autonomously carry out influence operations—planning and executing a multi-step campaign end-to-end without human prompting. With safeguards in place, its latest models refused nearly every task, the company said.

On the question of political neutrality, the company runs evaluations before each model launch to measure how consistently and impartially Claude engages with prompts expressing views from across the political spectrum. Opus 4.7 and Sonnet 4.6 scored 95% and 96%, respectively.

For users seeking voting information, Claude will surface an election banner directing them to TurboVote, a nonpartisan resource from Democracy Works that provides reliable, real-time information about voter registration, polling locations, election dates, and ballot details. A similar banner is planned for Brazil's elections later this year.

Anthropic said it plans to continue monitoring its systems and refining its defenses as the election cycle progresses. Decrypt reached out to Anthropic for comment on the findings, but did not immediately receive a response.

Daily Debrief Newsletter