Why Anthropic, and what draws you to AI safety?

[INTERVIEWER] Why Anthropic, and what draws you to AI safety? "Why Anthropic, and what draws you to AI safety?" Generic passion loses this question every single time. "I care about beneficial AI" is what everyone says, and it lands like wet paper. What they want is a specific, honest reason tied to what Anthropic actually does, plus a personal thread that makes the safety interest believable instead of rehearsed.

Reference their real work, then connect it to something you've genuinely done or thought about. The interviewers are checking whether your interest in safety is real and informed, or just a costume you put on for the occasion. Safety washing is easy to spot. The tell is specificity. Someone who actually cares can name the exact thing that made them care.

I'll give you the three part structure, a full worked example, and the follow up questions they use to test your motivation. Three parts, and keep every one of them honest. Part one: the specific "why them." About ninety seconds. Name concrete things about Anthropic, not just the mission statement. Real anchors you can reference: Constitutional AI, which is training a model against a written set of principles rather than only human feedback.

Their public work on interpretability, trying to understand what's actually happening inside the model, not just its outputs. The Responsible Scaling Policy, which ties capability releases to safety thresholds. And their eval and red teaming culture. Pick one or two you genuinely find interesting, and say why. Depth on one beats name dropping five. Part two: the genuine "why safety." About two minutes.

Your real thread. Where did you first care that a model being wrong, or confident, or misused, actually mattered? Tie it to something you did. A time you caught a model failing in a way that would've hurt a user. A decision where you slowed a launch to add a guardrail. Safety interest is credible when it comes from direct experience, not from general feelings or abstract philosophy.

Part three: why you, as a PM, here. About ninety seconds. What you bring. Shipping useful products while holding the safety line. Being the person who insists on the eval. Translating research caution into product decisions. And crucially, show you see safety and usefulness as one job, not two opposites. Here's how that sounds joined up. "Two specific things pull me here.

First, Constitutional AI. The idea that you can train a model against a written, inspectable set of principles, instead of only opaque human preference data, is exactly the kind of thing I want to build products on. Because as a PM, I can actually reason about that behaviour and explain it to a stakeholder. Second, the Responsible Scaling Policy, because tying releases to safety thresholds is the same discipline I've tried to bring at a much smaller scale.

On why safety, it's honestly not abstract for me. On my last team we were about to launch an automated customer support routing tool. I noticed the model was confidently misclassifying urgent escalation requests as low priority when users used certain dialects. I slowed the launch by two weeks to add a confidence threshold guardrail and a human fallback loop.

That taught me something I haven't let go of. The gap between a demo working and a system being safe to put in front of real people is where most of the harm lives. Closing that gap is genuinely the part of the job I enjoy most. What I bring is the product side of that. I ship things people actually use, and I'm the PM who won't let the eval get cut for the deadline.

Anthropic is one of the few places where that instinct is the point, not the friction." Notice it never praises the company back at them. It names two real things, gives one real moment, and connects both to how you work. Here's what makes them lean in. First, specific references to their actual work, with a real reason you find it interesting, not a Wikipedia summary.

Second, a safety motivation grounded in something you did, a real moment, not a mission statement echo. And third, framing safety and usefulness as the same job. That's how they see it internally, so hearing it back from you tells them you'd fit the way the team actually operates. Now the ways people lose this. The first is generic "I care about beneficial AI" with nothing specific to Anthropic in it.

If your answer would work word for word for any AI lab, it's not an answer. The second trap is praising the company back at them with buzzwords instead of one honest, concrete thread. Flattery reads as filler. And the third is treating safety as the opposite of shipping, which comes across as either naive or performative, because the whole company exists to prove those two things go together.

Expect the follow up. They'll ask "what would you actually disagree with us on?" or "where would you push back?" This is a good sign, not a trap. Have one honest, considered thing ready, because a candidate who agrees with everything is a candidate who hasn't thought about it. And they may ask "give me an example where you chose safety over speed."

That's why your real moment matters. The delayed launch story answers it directly. So the whole picture: one or two specific things about their work you genuinely find interesting, and why. One real moment that made you care about safety, tied to something you actually did. And what you bring as a PM who treats safety and usefulness as one job.

Keep all three honest, because the entire question is a sincerity test, and specificity is the proof. One line to carry in: one specific thing about their work you genuinely find interesting, plus one real moment that made you care about safety. Specificity is sincerity here, because anything vague just reads as a costume.