Summary: On October 14, 2026, Anthropic published its preparation plan for the US midterms, covering how people actually use Claude for politics, what its bias and misuse testing found, and where its defenses fell short. For anyone shipping AI products in a trust-sensitive domain, it’s a useful worked example of how to measure demand, expose gaps honestly, and patch them before a deadline.
Start with the usage data, not the fear
The most grounded part of Anthropic’s post is the measurement. Using its Anthropic Insights tooling, the company analyzed roughly 250,000 US Claude.ai conversations from September 21 to 27, 2026, with a privacy-preserving classification method. About 5.75% of US conversations touch on politics or policy; about 0.51% are specifically about elections or voting. Requests asking Claude whom to vote for showed up in around 0.01% of conversations, and policy-violating requests in just 0.002%.
That last number matters. It’s a reminder that the scariest misuse scenario is often vanishingly rare in real traffic, while the boring informational use case — “where candidates stand on X” — dominates. If you’re building guardrails, measure the actual distribution before you decide what to defend against hardest.
Connect answers to authoritative sources
Instead of trying to make the model a voting authority, Anthropic is piping in data from Democracy Works, the nonprofit behind TurboVote. Users in the US asking where to vote or what ID to bring will get state-specific answers drawn from what Democracy Works collects from state and local election sources, directly in the conversation. Claude will also surface official election-night results as they update.
The pattern generalizes: for high-stakes factual categories, retrieval from the authoritative source beats hoping the model knows. It’s the same instinct behind grounding search agents and structured data pipelines — push the risk to a pipeline you can audit rather than to weights you can’t.
Red-teaming found a real gap: death by a thousand harmless steps
Anthropic partnered with the Institute for Strategic Dialogue and the McCain Institute to pressure-test safeguards across the US, Israel, and Brazil. The headline finding: Claude reliably refused a single harmful request, but could miss a malicious campaign when it was split into many individually innocuous tasks — audience research, building a news site, drafting content, planning distribution.
That’s the failure mode most teams don’t test. Single-turn safety evals pass; multi-step, multi-request abuse slips through. Anthropic’s response was to add classifiers targeting novel and multi-step misuse, including fake social media accounts and pseudo-independent news sites, plus detection of financially motivated election schemes.
Bias testing that names its own shortcomings
The company also ran external evaluations — the Future of Free Speech at Vanderbilt, the Foundation for American Innovation, and the Collective Intelligence Project. Its newest models score 99% on its political evenhandedness evaluation, but the outside groups found uneven hedging: Claude Sonnet 4.5 was more likely to soften right-leaning arguments, and Claude sometimes inserted unsolicited caveats or counterarguments. Anthropic says it’s fixing those specific behaviors.
This is worth internalizing. It echoes what I noted in the earlier writeup on Anthropic’s self-replication tests — publishing what the safety work did not achieve is what makes the rest of the claims credible. A 99% score means little without the 1% spelled out.
The takeaway
If your product touches anything contentious, Anthropic’s sequence is a reasonable template: measure real usage, ground answers in authoritative data, red-team multi-request abuse paths, and publish the gaps you found. The limitation is that this playbook depends on scale and dedicated threat-intel teams most builders don’t have — but the measurement-first ordering costs almost nothing and is the part worth copying first.