Intel / article

Election Safeguards Update

Source: anthropic.com
Author: Anthropic News
Date: Published Apr 28, 2026

Context

Anthropic's April 2026 election update tested whether Claude looked for fresh information when asked about the US midterms. Using over 200 distinct prompts, each in three variations (over 600 in total), covering candidates, voting procedures, polling, election dates and key races, Opus 4.7 triggered web search 92% of the time and Sonnet 4.6 did so 95% of the time, according to the published account.

Web search matters because Claude has a and will not automatically know about developments such as candidate announcements, and Anthropic notes it can relay current information when search is enabled. The percentages measure how often the models went looking for current material on these prompts. They do not measure whether the resulting answers were correct, which is a separate question; Anthropic itself advises verifying important information with official sources.

Terms in this piece · Glossary

knowledge cutoff — The date after which a model saw no training data — everything later has to reach it through search, tools, or the context window.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.

Read the source www.anthropic.com