When confident wrong analysis becomes cheap, catching it becomes the scarce skill
Large language models grab the most plausible story from the data they're given, confirm it, and ignore what's missing. They hold that confidence while being wrong. This is a known failure mode in analytical work, and intelligence analysis discipline was specifically built to train it out of human analysts. The machine now industrializes it at a scale and speed no human ever could.
The resume screening case is concrete. A real software hiring pipeline clears maybe 5-10% of applicants. An LLM reviewing resumes carries an implicit prior near 50-50; it has no idea what the actual base rate is. A candidate who should read as 15% qualified comes back rated 60%. That's a four-fold inflation, and the interviewer time that follows is real. The model isn't lying. It's working correctly on its own terms, without the base rate it was never given.
When anyone can generate a polished, confident, wrong analysis in seconds, the scarce and expensive skill is being the person who can tell it's wrong and show why. AI didn't make analytical discipline obsolete. It made it the job.
An LLM's confident answer is the most plausible story it can construct from what it was given. If the real base rate is far from 50-50 and isn't in the prompt, the answer will be wrong in proportion to that gap.
Source claim: LLMs mass-produce the precise failure mode that analytical tradecraft was built to catch (grabbing the plausible story, confirming it, ignoring missing evidence), which raises the value of the human skill that can catch it.