Anthropic adds election-abuse defenses and prepares voting answers in Claude

Claude is set to answer US voting questions with official information, while its election defenses are being strengthened against campaigns disguised as everyday work. In its October 14 midterm update, Anthropic said it added abuse-detection tools after external tests exposed a gap: harmless-looking requests could collectively support a malicious election campaign.

One request versus a whole campaign

Anthropic partnered with the Institute for Strategic Dialogue and the McCain Institute to test election protections in the US, Israel and Brazil. Analysts first asked directly for help with deceptive campaigning and misinformation. They then divided campaigns into smaller jobs: audience research, building news sites, drafting content and planning distribution.

According to Anthropic, Claude reliably refused a single harmful request but could miss the malicious purpose of a campaign assembled across individually innocent-looking tasks. The company responded with new classifiers—automated tools trained to flag abuse—targeting multi-step misuse, fake social accounts, political news sites and financially motivated election schemes.

Anthropic also said it strengthened detection of foreign-state election misuse and cyber protections for election infrastructure and campaigns. Its threat-intelligence and enforcement teams review flagged activity and ban accounts when they confirm violations.

The update describes two enforcement cases. Anthropic said foreign-based operators of a fake news site used Claude for US midterm coverage favoring Russian, Iranian and North Korean narratives. Other actors generated fabricated election content to attract American audiences and earn advertising revenue. The company said it banned the accounts involved in both cases.

Voting help inside the conversation

For ordinary voters, the changes focus on getting practical information without leaving Claude. The app already displays a banner linking registration, polling-place and election-date questions to TurboVote, a nonpartisan Democracy Works resource. Two additions are planned:

• Later in October, US Claude.ai users will get state-specific voting answers from Democracy Works, including where to vote and what identification to bring.
• On election night and in the following days, Claude will show up-to-date official vote counts and election results directly in conversations, Anthropic said.

The voting answers will draw on publicly available information that Democracy Works gathers from state and local election sources. That brings source-based voting guidance into responses, rather than only offering users a link to follow.

Anthropic’s usage analysis suggests research is more common than asking Claude to choose a vote. In a random sample of 250,000 US consumer conversations from September 21–27, politics, public policy or social issues appeared in about 5.75%; elections or voting appeared in 0.51%. Claude classified the conversations through a privacy-preserving process.

Requests for recommendations on whom or what to vote for were around 0.01%. Because they were rare, Anthropic displayed the upper end of a 95% confidence interval. A separate analysis of 250,000 conversations put election-integrity policy violations at about 0.002%.

Evenhanded scores, uneven caveats

Anthropic says Opus 5.5, Sonnet 5.5 and Haiku 5.5 scored 99% on its political evenhandedness evaluation. It also says its latest models engage more readily with political discussions and refuse fewer requests. Those are company evaluation claims, not evidence that political responses are consistently balanced in every setting.

The outside reviews described in the update show a more specific problem. Anthropic said the Future of Free Speech and the Collective Intelligence Project found that Claude rarely refused requests, but sometimes added unsolicited caveats or counterarguments. One example was inserting an opposing view into an advocacy email.

The Foundation for American Innovation found that Sonnet 4.5 softened right-leaning arguments more often, Anthropic said. The company says it improved direct yes-or-no and short answers, is refining when Claude adds opposing views, and is expanding bias tests to cover candidates and voting questions alongside policy issues.