Anthropic
Blogcast

Alex: Welcome back to Anthropic Blogcast, the show where we turn Anthropic's latest posts into a listen, so you can keep up without opening a dozen tabs.

Jordan: Today we are talking about elections, because Anthropic has published an update on how it is preparing Claude for the United States midterm elections this year, and the post is really about three jobs at once, giving voters good information, staying evenhanded, and stopping misuse.

Alex: That framing helps, because this is not a product launch. Anthropic says it shared an earlier version of these safeguards this year, and this post explains what it has built since then as Election Day gets closer.

Jordan: Let us start with how people actually use Claude around politics, because Anthropic measured that first. It looked at a sample of about two hundred and fifty thousand conversations from one week in late September, using a privacy preserving method to classify what the conversations were about.

Alex: And the picture is more ordinary than a lot of people might expect. Conversations touching broadly on politics, public policy, or social issues made up a little under six percent of conversations in the United States on the consumer app. Conversations specifically about elections or voting were about half of one percent.

Jordan: The detail that stood out to me is how rarely people ask Claude to decide for them. Requests asking whom or what to vote for showed up in only about one hundredth of one percent of conversations. Much more often, people asked where candidates stand on an issue, or asked for help understanding news, research, or government services, so they could reach their own conclusion.

Alex: And requests that broke the election integrity rules were rarer still. Anthropic says the typical political conversation is informational, things like making sense of world events, researching candidates, drafting documents, or weighing arguments for and against a policy.

Jordan: Which leads to the second job, connecting people to trusted sources instead of asking them to trust the model alone. When someone asks about voter registration, polling places, or election dates, Claude will show a banner linking to a nonpartisan voting resource, and that part has been in place since the spring.

Alex: The new piece is that later this month, for users in the United States, Claude will start pulling official voting information directly into its answers. So if you ask where to vote or what identification to bring, the answer can be specific to your state, drawn from information collected from state and local election sources, without you having to leave the conversation.

Jordan: Anthropic also says Claude will show up to date results on election night and in the days after, so a question about current vote counts points to the latest official numbers rather than a stale summary the model might produce on its own.

Alex: That is a sensible division of labor. Let the model explain and summarize, but let official sources supply the facts that change by state and by hour, like polling locations and results.

Jordan: Now the third big theme before we get to misuse, political bias. Anthropic says it tests every model before launch for political bias, and it describes its latest models as its most evenhanded yet, scoring ninety-nine percent on its own evenhandedness evaluation, and refusing fewer political requests than earlier versions did.

Alex: Anthropic did not stop at its own test, though. It asked three independent groups to probe the same question from different angles. One think tank tested hundreds of requests on contentious topics, another group compared responses to matched prompts from the left and the right, and a third group surveyed more than two thousand people across three countries to help build shared principles for judging responses.

Jordan: And to Anthropic's credit, the post reports where those outside reviews found problems, not just praise. The reviewers found that Claude rarely refused, but sometimes added caveats or counterarguments nobody asked for, for example slipping an opposing view into an advocacy email someone had asked it to draft.

Alex: Another review found that an earlier model did not always soften its language evenly, and was more likely to soften arguments for right leaning positions. Anthropic says it is responding by improving answers to direct yes or no questions and to requests for short answers, by being more consistent about when it adds caveats, and by testing more carefully for uneven hedging.

Jordan: That is a useful reminder that evenhandedness is not just about refusing to answer. It is also about tone, about whether the model argues back with one side more than the other, and about whether a short answer stays balanced when there is less room to explain.

Alex: Let us turn to misuse, because that is where the rules get concrete. Anthropic's usage policy bans using Claude to deceive voters or disrupt elections, whether that means spreading false information about candidates or about how to vote, impersonating election officials, or trying to suppress turnout.

Jordan: At the same time, Anthropic is clear that campaigns do a lot of legitimate work, and it recently published a guide to help them see the line. Drafting press statements and fundraising emails, preparing for debates, analyzing turnout data, and building voter outreach materials are all presented as normal uses.

Alex: The prohibited side is equally concrete. Creating fake social media profiles that pretend to be real people, mass producing comments to inflate a campaign's engagement, or writing for news sites that falsely claim to be independent all cross the line.

Jordan: Then there is the testing, and this is where the post gets most interesting for builders. Anthropic worked with two independent research organizations to pressure test its defenses in the United States and in two other countries holding major elections this year. Testers started with direct harmful requests, and Claude refused those reliably.

Alex: The weakness appeared when a harmful campaign was broken into many small steps that each looked harmless on their own, things like audience research, building a news site, drafting content, and planning distribution. A single request can look fine while the whole operation is deceptive, and that is exactly the pattern automated safeguards can miss.

Jordan: Anthropic says it responded by adding classifiers aimed at that multi step pattern, including attempts to build fake accounts and fake news sites for political influence, and financially motivated election schemes. It also says it strengthened detection of foreign state actors and its cyber protections for election infrastructure and campaigns.

Alex: And the threat findings give that work a real example. Anthropic describes foreign based operators behind a fake news site who used Claude to produce midterm coverage favoring the narratives of several foreign governments, and separate actors who produced fabricated news to attract an American audience and earn advertising money. Anthropic says it banned the accounts in those cases and improved its detection.

Jordan: So if we step back, what should a listener take away? First, most people are not using Claude to be told how to vote. They are using it to understand issues and candidates, which makes accuracy and sourcing the main job. Second, Anthropic is moving official voting information and election night results into the conversation itself, rather than leaving the model to answer those questions from memory.

Alex: I would add two more. Third, the bias work is now being checked by outsiders, and the flaws they found were about subtle things like unsolicited caveats and uneven softening, not just refusals. And fourth, the hardest misuse problem is no longer a single bad request. It is a campaign spread across many ordinary looking requests, and that is where Anthropic says it is now focusing its classifiers.

Jordan: That last point travels well beyond elections, too. Any safeguard that only judges one request at a time will struggle with work that is harmful only when you see the whole pattern, and a lot of agent safety work is heading in that same direction.

Alex: Anthropic closes by saying it plans to share a fuller look at how people used Claude across the election cycle in the coming months, and that this work extends to elections around the world, not just the United States. We will watch for whether that follow up shows the safeguards held up in practice.

Jordan: Thanks for listening to Anthropic Blogcast. We will be back with the next Anthropic post as soon as there is a new one worth your ears. Until then, take care, and keep building.