Is Claude Becoming the Government's Voice?!

Claude may soon answer U.S. voters' questions directly, with state-specific information pulled from election sources. Anthropic says the direct-answer feature isn't live yet. The same week, Tesla's next-generation Roadster lifted off the ground and rotated in a demonstration, and half the internet called it a flying car.

Two very different stories, one question underneath both of them, and it's the one Oscar and I kept coming back to in Episode 28: what does an announcement or a demo actually prove? Usually less than the headline says. Sometimes more than anyone's paying attention to.

Signal or Noise

Same filter as every week: is this something you should act on, or noise dressed up as news?

Claude Voting Info

Anthropic plans to have Claude answer voter-information questions with state-specific details, working with Democracy Works. On its face this is logistics: where to vote, what to bring, when the deadline is. The part I care about is everything around the answer. Where did it come from, does it match where you live, and how fresh is it? An election answer that's right for the wrong state, or right as of last month, is worse than no answer. I have a lot more to say about this one, so it became my unpopular opinion this week. It's the whole section below.

Tesla Roadster Jump & Rotate

The Roadster briefly lifted and rotated in a demonstration. It's a striking clip. It is not a flying car. A prototype doing something once on camera is a very different claim from a delivered, road-approved passenger vehicle, and the gap between those two is where most of the actual engineering lives. This is the cleanest example of the episode's theme: grade the demo for what it shows, not for what it implies.

Anthropic Warns of AI Self-Replication

Anthropic published red-team experiments on whether models can copy themselves. The headlines want this to be an AI worm loose on the internet. It isn't. What the report raises is a security question: if a model can do this under test conditions, what controls do you need around the systems you deploy? That's a useful question for anyone running agents with real credentials, and it's a much less exciting one than the headlines.

Bank of America AI Vulnerability Remediation

Bank of America reportedly brought in about 1,000 contractors for AI-linked vulnerability remediation, alongside roughly 3,000 existing personnel. It's tempting to read that as "AI found bugs faster than humans could fix them." The figures don't prove that. They also don't prove every contractor worked only on Mythos findings, and the reported remediation costs included separate compliance work. What it does show is the part nobody puts in the launch post: finding problems is cheap now, and fixing them still costs headcount.

Anthropic Reports Decline in Junior Hires

Anthropic reported a decline in the share of junior technical hires. These are company-specific figures, not an industry-wide forecast, and I'd be careful about anyone who turns one company's hiring mix into a prediction about everyone's career. It's still worth watching. If the companies building the tools are hiring fewer juniors, the rest of us should be asking how people learn the job in the first place.

Ship It or Skip It

This week's segment stayed close to home: two ways to check the code an AI writes for you.

1. Mutation testing. Deliberately break your code in small ways and see whether your tests notice. If AI is writing both the code and the tests, a green test suite proves less than it used to. Mutation testing is a tougher check on whether those tests actually test anything. Stryker Mutator is the tool we pointed at, and the MuTAP research looks at using models to strengthen tests this way.
2. AI code review. A second model pass over generated changes, with a human still responsible for the merge decision. That last clause is the whole point. A reviewer that can't be held accountable is a suggestion engine, not a review. We didn't endorse a vendor or claim a performance result. The idea is the pattern, not the product.

My unpopular opinion: point, don't pronounce

An AI assistant should never speak in the government's voice. The moment Claude answers civic questions directly from "official" sources instead of pointing you to them, it stops being a tool and starts being an announcer, and I hate it.

It's only half a joke. Not because the information is wrong. Where to vote, what ID to bring, that's logistics. Nobody's arguing about where the gym is.

I hate it because of the pattern being built. The lab decides which source counts as official, and then the model repeats it as fact, in its own voice.

Think about what that does to the answer. A link says, "Here's who said this. Go check." A confident sentence from an AI says, "This is true." The source disappears. The provenance disappears. And the user has no way to tell when that official source is outdated, or wrong, or has been captured by somebody with an agenda.

Patterns like this don't stay in their lane. Once there's an "official source" pipeline wired straight into the model, every government, every agency, every pressure group on earth is going to want their source on that list. And not just for polling places.

We've seen this movie before. For a good chunk of the printing press's history, governments licensed what could be printed. England didn't let its press-licensing law lapse until 1695. Official information has always been right up until the moment it wasn't.

The danger isn't that Claude tells you the wrong polling place. The danger is that we're training hundreds of millions of people to accept an AI speaking in the government's voice, and to stop asking who's actually talking.

So here's what I'd ship instead: point, don't pronounce. Give me the election office. Give me the link. Give me the date it was last updated. And let the human be the human in the loop.

Oscar's take

Oscar's unpopular opinion came at the same problem from a different direction: he doesn't want opinions from his AI at all. Not on politics, not on movements, not on anything. A model should be a logic machine that gives you the information you need and offers no personal opinions, because it shouldn't have any.

We're closer than it sounds. He doesn't want the model to have a view; I don't want it borrowing anyone else's view and saying it in its own voice. Either way, the user should be the one deciding what's true.

Where I landed

The thread through this whole episode is the gap between what something looks like and what it proves:

• A demo proves the thing happened once, on camera.
• A report on self-replication proves a capability under test, not a threat in the wild.
• A contractor headcount proves remediation is expensive, not why.
• A confident answer proves nothing at all unless you can see where it came from.

That last one is the one I'd hold onto. The more fluent these systems get, the more the source matters, and the easier it is to lose.

Watch or listen to the full conversation below, and tell me which side you're on: should your AI answer civic questions, or just point you to who can?

---

Human in the Loop — Watch on YouTube · Listen on Spotify · Listen on Apple Podcasts

Every episode — full show notes, sources, and the archive — lives at podcast.vallyseed.com. The show is produced by VallySeed, where Oscar and I help organizations design, build, and deploy AI systems that create measurable competitive advantage.