AI hallucination: a field guide from someone who ships

After enough time shipping AI features, you develop a nose for where hallucinations live. They're not evenly distributed. They cluster exactly where precision matters most, which is why they're dangerous.

Where they cluster

Numbers. Statistics, prices, dates, quantities. The model produces numerically plausible text, which is different from true text.

Citations and names. Titles that almost exist, authors who almost wrote that, links that almost resolve. If you've ever chased a reference and found nothing, you know this one.

Specifics inside general fluency. The most dangerous hallucinations ride inside an otherwise correct answer. Ninety accurate sentences and one invented API parameter. Confidence doesn't drop where accuracy does.

Anything with a version. Models trained on yesterday's docs will answer today's questions with yesterday's API, without hesitation.

What actually helps

Grounding: give the model the source text and instruct it to answer only from what's present, and to say "not in the document" otherwise. Validation: schema-check every structured output. Friction: make verification cheaper than error — for anything consequential, the human step isn't optional, it's the design.

The goal was never zero hallucination. It's systems that know where they're blind, and interfaces that make the checking easy.