Korva is a scroll of the internet’s best ideas.
Don't delegate decisions to a model
Drawing a boundary
Phones are a good comparison: they can fry your brain, but you can use one all day if you learn where your own boundary is, such as when it's fine to scroll through reels and when it isn't, like in a meeting. For work, the boundary sits between judgment and decision-making on one side and execution on the other.
Trust and check
Models make it look like they're doing the right thing, so everything still needs double-checking with a little paranoia. A person who delegates their actual job, the decision-making, to a model loses that ability.
Judgment in product work
The worry is practical for product teams. Running a lot of experiments only helps if each one actually reveals information instead of being fudged, and with AI tools it is easy to offload a lot of thinking and judgment. The aim is for people to keep real judgment and know what they're doing instead of regurgitating what comes out of Claude.
Why does relying on tools weaken skills?
A 2011 study in Science by Betsy Sparrow, Jenny Liu, and Daniel Wegner found that people remember information less well when they expect to be able to look it up later. A 2020 study in Scientific Reports by Louisa Dahmani and Veronique Bohbot linked habitual GPS use to poorer spatial memory. Psychologists call the handing of mental work to an outside aid cognitive offloading.
Why do models sound right when wrong?
Language models generate text by predicting likely next words, so fluent phrasing comes easily whether or not the content is correct. Training on human feedback can also reward answers that raters find convincing. Anthropic researchers published a 2023 study showing that models trained this way sometimes favor answers matching a user's stated views over accurate ones.
Why do people over-trust automated suggestions?
Researchers call it automation bias. A 2010 review in the journal Human Factors by Raja Parasuraman and Dietrich Manzey covered studies in aviation, medicine, and other fields where people accepted automated advice even when it was wrong. Monitoring a system that is usually right also dulls attention over time.
More from 20VC with Harry Stebbings
