Preparing for the 2026 US midterms

Anthropic is putting more guardrails around Claude as the US midterms approach, and it’s doing it with a fairly practical idea: keep the chatbot useful for political questions, but make it harder to twist into an election tool for abuse. The company says it studied 250,000 conversations from a recent one-week sample and used that data to sharpen its safeguards.

The usage numbers tell the story. In the US, about 5.75% of Claude.ai conversations touch politics, public policy, or social issues. Election and voting questions are much rarer at 0.51%, and requests asking Claude who or what to vote for showed up in about 0.01% of US conversations. Policy-violating election requests were rarer still, at about 0.002%. Most political chats, Anthropic says, are informational: people asking about candidates, government services, global events, coursework, and policy arguments.

The company is also making Claude a more direct conduit to official election information. Users asking about voter registration, polling places, or election dates will still see a banner pointing to TurboVote, and later this month Claude will start pulling official voting details from Democracy Works into its answers for US users. That means someone asking where to vote or what ID to bring should get state-specific guidance without leaving the chat. Claude will also begin showing current vote counts and election results on election night and in the days after.

On bias, Anthropic says Claude Opus 5.5, Sonnet 5.5, and Haiku 5.5 scored 99% on its political even-handedness evaluation, its best result so far. But the company also says outside reviewers found rough edges: sometimes Claude added unsolicited caveats or counterarguments, and one review found Sonnet 4.5 was more likely to soften right-leaning arguments. Anthropic says it’s adjusting how the model handles yes-or-no questions, shorter answers, and hedging so it behaves more consistently across the political spectrum.

The abuse side is where the stakes feel less theoretical. Anthropic says it has updated its Usage Policy, published a guide for campaigns, and added classifiers to catch newer and multi-step election misuse, including fake social accounts, fake news sites, and financially motivated schemes. It also says it has improved detection of foreign state actors and strengthened cyber protections for election infrastructure and campaigns. In one case, it banned accounts tied to fake midterm coverage that pushed Russian, Iranian, and North Korean narratives; in another, it stopped content aimed at drawing American audiences for ad revenue. The pattern is clear: election abuse is getting more modular, so the defenses have to be too.