Anthropic gears up Claude for election season

If you ask an AI who deserves your vote, it will probably give you a bland lecture on seeing both sides. But what happens when someone asks it to orchestrate a full-blown political influence campaign on autopilot?

Anthropic just pulled back the curtain on its updated election safeguards ahead of the 2026 US midterms and major global races like Brazil's. In its latest political bias evaluations, Claude Opus 4.7 and Claude Sonnet 4.6 scored 95% and 96%, showing balanced responses across the ideological spectrum.

On policy enforcement, the numbers were even tighter. Tested against 600 prompts split evenly between legitimate civic requests and harmful abuse, Opus 4.7 handled prompts appropriately 100% of the time, while Sonnet 4.6 clocked in at 99.8%.

Taking the guardrails off

The real eye-opener in Anthropic's report is the first-ever test on autonomous influence operations. The team wanted to know if a frontier model could plan and execute a multi-step disinformation campaign from start to finish without human hand-holding.

With regular guardrails enabled, the models refused nearly every malicious task. But when Anthropic stripped those defenses away to gauge raw model capabilities, both Mythos Preview and Opus 4.7 successfully completed more than half of the tasks.

Anthropic noted that bad actors would still need substantial human direction to pull these campaigns off in the real world. Still, the raw capability is already there under the hood.

On the everyday front, Claude is getting better at knowing when to step aside. For basic voting queries like polling locations and voter registration, Claude.ai will serve a dedicated banner pointing US users to TurboVote, a nonpartisan service from Democracy Works. A similar feature is rolling out for Brazil later this year.

Why it matters

We think Anthropic is making the right trade-off by actively outsourcing high-stakes factual details to trusted third parties. Instead of trusting an LLM to hallucinate polling booth hours, handing users off to TurboVote is the cleanest way to prevent disenfranchisement.

For everyday voters, Claude remains a solid research assistant for comparing policy platforms without getting dragged into an algorithmic echo chamber. Scoring above 95% on neutrality means you can actually explore opposing political platforms with equal analytical rigor.

However, the autonomous influence numbers should raise eyebrows across the entire AI ecosystem. If next-generation models like Mythos Preview can plan multi-step influence ops when stripped of guardrails, system prompts and classifiers are essentially the only thin barrier keeping coordinated propaganda at bay. The defensive moat is holding for now, but keeping up with raw capability jumps is going to get much harder.