The AI expectations gap

Edmund Anthropic’s staff think AI is going to eat way more of their own jobs by 2030 than the public does. That’s the whole crack in the floorboards here.

Geffen Yeah, and the weird part is they don’t seem wildly split on what AI can do today. They’re split on how fast it gets from demo to day-to-day use, which is the annoying part people keep hand-waving past.

Edmund Right. The article has that clean number too: ninety-one percent of staff think AI can make a Nobel-level discovery by 2030, versus thirty-nine percent of US adults. That’s not a tiny disagreement.

Geffen No, that’s a canyon. And I like that they didn’t just ask some vague 'is AI good' question, because that would’ve been useless soup.

Edmund They also ask about actual work, which is where it gets more interesting. Over half the staff think AI can do at least eighty percent of their current work by 2030, while only five percent of US workers say that about their own jobs.

Geffen And the public isn’t clueless, exactly. People who already use AI at work are more aggressive than non-users, but they still land way below the staff view.

Edmund That’s the part I care about. If you’ve used the tools enough to get the little workflow wins, you start moving the forecast up. But Anthropic staff are living inside the machine every day, so of course they’re seeing the ceiling differently.

Geffen Sure, but I wouldn’t over-read the intimacy. Being close to the thing can make you better calibrated, or it can make you overfit the local curve and think the rest of the world is about to snap into place.

Edmund Fair. The article’s stronger move is that it separates capability from adoption. The public thinks the big blockers are human preference, privacy, reliability, and regulations, not just raw tech limits.

Geffen That’s probably right. Customers liking humans is not a small variable, and privacy plus security are real brakes, especially in work settings where the task might be doable but still not shippable.

Edmund Exactly. And the staff expect much faster diffusion too. They think AI will be used for sixty-five percent of the tasks it could be used for, while the public says forty-two percent.

Geffen That’s the adoption gap in one line. Capability can be there, and the actual usage still lags because somebody has to trust it, wire it in, and live with the checking cost.

Edmund The time-savings piece is the most product-shaped part of the whole note. Staff say a task that AI is well suited for takes half the time or less for eighty-five percent of them, and fifty-nine percent say it takes a quarter of the time or less.

Geffen Yeah, that’s huge. The public sample is much more modest, and a noticeable chunk even says it takes as long or longer, which tells you the checking burden is still very real for them.

Edmund And honestly, that’s where I think the economics lives. If the tool saves ten percent, it’s a nice helper. If it saves seventy-five percent on a task people actually do all day, then the workflow changes shape.

Geffen Right, but the paper is careful enough not to claim the productivity number is magic. A lot of this is self-report, and staff may be describing their own internal tooling environment, which is a pretty special case.

Edmund Yeah. Not the median company basement. But I don’t think that kills the article. It just means the staff answers are a view from inside a place that’s already reorganized around the tools.

Geffen And that matters because the article’s labor-market point is subtler than 'AI means unemployment, period.' It says people who expect AI to do more of their own work also expect higher unemployment, and both groups reason that way in basically the same shape.

Edmund That’s the part that actually holds up to me. They may disagree on the slope of the curve, but once you hold automation constant, they’re pretty aligned on the labor-market intuition.

Geffen Which is a nice result, because it suggests the fight is not some mystical disagreement about economics. It’s mostly about how quickly the same automation story becomes real in actual jobs.

Edmund And that’s why this matters for people building products. If you’re selling AI into work, the question isn’t just 'can it do the task.' It’s whether the user believes the checking, trust, and handoff costs are low enough to let it in.

Geffen I’d put a small bet on the broader public staying more cautious than the staff through the end of next year, just because lived workflow friction is slower to disappear than model demos. Nothing dramatic, just the gap staying visible.

Edmund Yeah, I buy that. Not because people are pessimists, but because reality keeps showing up with a clipboard. Anyway, this is a nice one for us to argue about in the same old kitchen-table way, which is apparently the brand now.

Geffen Unfortunately, yes. Come on, Edmund, we’ve become the sort of show that reads survey tables for fun.

Edmund And yet here you are, fully engaged. All right, let’s leave the table standing and call it there.