← All explainers
AI in 3 · 3 min read

Hallucination is not a bug you can patch

Why fluent invention is intrinsic to the method, and what to design instead of a fix.

AI in 3Foundational

These models predict plausible continuations. A confident wrong answer and a confident right answer are produced by exactly the same mechanism — there is no internal flag separating them.

So the product question is never "how do we stop it". It is "what happens when it is wrong, and who notices". That answer belongs in the spec, next to the success metric.

Grounding with retrieval, citing sources the user can check, showing confidence honestly, and putting a human in the loop for irreversible actions all reduce harm. None of them reduce it to zero, and a design that assumes zero is the actual defect.

The shape of it

  1. 01

    Ground

    Answer from retrieved sources and cite them inline.

  2. 02

    Constrain

    Structured output and validation beat prose for anything downstream.

  3. 03

    Verify

    Second-pass check for high-stakes claims — numbers, names, policies.

  4. 04

    Contain

    Human approval before anything irreversible or externally visible.

Want this applied to your situation?

Sessions are direct and specific — you leave with a decision, not a reading list.

Book a session
More in AI in 3