Geek Out Time: Can a Model Detect When It’s About to Hallucinate?

Hallucinations are usually discussed after the fact. The answer is wrong. The reference doesn’t exist. A confident explanation collapses the moment a human checks it. At that point, we talk about better prompts, stricter retrieval, guardrails, or simply using a bigger model. Before a model hallucinates, is there anything observable that already looks risky?

Not philosophically. Not in hindsight. But inside the model, while it is in the middle of deciding what to say next. This Geek Out is a small experiment around that question.

A setup on Google Colab

I kept the setup intentionally lightweight. The goal wasn’t statistical coverage or benchmark performance, but observability.

This is an excerpt — the full article continues on Medium.