Someone in a meeting says something right about a part of the system they don’t own. A precise point, with the constraint that matters sitting inside it. A year ago they’d have stayed quiet.
You’ve probably watched this happen. You may have been that person.
The question only shows up later, never in the room. Did they know that, or did they read it twenty minutes ago?
Here’s what I keep circling back to. There’s a gap now between sounding right and having checked, and from the outside you can’t see it. You can’t hear it either. The answer arrives already shaped like a good answer, and the shape is the part that used to be expensive.
That’s worth being nervous about. It’s also the best thing that’s happened to my meetings in years.
Nobody gives the tool back, for one. Someone tried to re-run a study on how much AI speeds developers up and had to abandon it. They couldn’t recruit enough people willing to work without the tool for the comparison.1 Not even at fifty dollars an hour.2
And when a company doesn’t hand the tool out, it goes on somebody’s personal credit card. Quietly, without asking. I’ve watched that happen often enough to stop calling it an anecdote, and nobody pays out of pocket for something that slows them down.
But the part I care about isn’t speed.
The floor came up. People who used to sit at the edge of certain conversations are inside them now. The New Stack got there before me on this: juniors will ask a model about code in ways they’d never ask the senior sitting two desks away.3
They haven’t become experts, and mistaking a plausible answer for a checked one is a real risk.
Still, if you’ve run a technical group you know what the silence of the people who don’t feel qualified actually costs you. That silence is rarely shyness. It’s a reasonable belief that you don’t have enough context to open your mouth, and that belief now dissolves earlier.
There’s a cost on the other side, and a Forrester analyst in the same piece names it. The fast answer can stand in for the slow conversation, and the slow conversation was also where the architectural context came from.4
So the same property does both things. Plausibility standing in for verification is the problem. Plausibility getting someone into a room they were locked out of is a gift. It’s one mechanism, seen from two corners of the room.
Which means squeezing the plausibility isn’t the move. Doing the checking afterwards is, on richer material than you had before. The colleague who says the right thing without fully knowing why it’s right is handing you something that needs verifying, the way everything always did. Before, it just wasn’t on the table.
I don’t think the judgment work goes away. I think it gets more expensive, because there’s more plausible material to judge — and whoever can do it is worth more than they were.
None of this is my idea, by the way. A group in Rome gave it a name last year. They call it epistemia: the feeling of knowing, without the work of judgment having happened.5 They got there by watching six models and a room of people rate the same sources. Everyone agreed on the verdicts. Nobody took the same road.6
Meanwhile, more people are talking.
Footnotes
-
METR (2026), We are Changing our Developer Productivity Experiment Design. ↩
-
The introverts’ edge: How AI is leveling the developer floor, The New Stack, May 2026. ↩
-
Andrew Cornwall (Forrester), quoted in The New Stack, May 2026. ↩
-
Loru, Nudo, Di Marco, Santirocchi, Atzeni, Cinelli, Cestari, Rossi-Arnaud & Quattrociocchi (2025), The simulation of judgment in LLMs, PNAS, doi:10.1073/pnas.2518443122. Link to the open preprint; the PNAS page is the canonical reference. ↩