---
id: con-verification
schema: concept/1
layer: garden
status: established
name: Verification
aka: ["fact-checking", "checking AI output", "double-checking", "source checking", "discernment"]
summary: >
  Checking what AI gives you against something real, such as the source, the data, an expert or your
  own knowledge, before you rely on it or pass it on.
first_seen: unknown
relations:
  builds_on: [con-hallucination]
  evolved_from: []
  related: [con-jagged-frontier, con-ai-writing-tells]
  parallels:
    - idea: "fact-checking"
      field: "journalism"
      note: "Every name, number and quote is checked against a primary source before publication."
    - idea: "the four-eyes principle"
      field: "finance"
      note: "Important transactions need a second person to confirm before they go through."
implementations:
  - { provider: perplexity, name: "Cited answers", resource: res-perplexity-evaluating-ai-high-stakes }
  - { provider: google, name: "Gemini double-check response" }
  - { provider: openai, name: "ChatGPT search with citations" }
competencies: [cmp-check-the-result, cmp-use-responsibly]
origin:
  kind: ai-proposed
  by: "seed agent 2026-10-16"
  rationale: "The rarest good habit in real use (fact-checking in under 1 in 10 conversations studied) and the one most worth building."
---

Verification is the habit of checking AI output before trusting it. It sounds obvious, yet in one large study of real conversations, people checked facts in fewer than one in ten. Polished outputs got checked even less.

You don't need to check everything equally. Match the effort to the stakes and to the kind of claim. Specific facts (numbers, names, dates, quotes, citations) need checking against a source. Reasoning needs a quick "does this follow?" Summaries of your own documents need a spot check against the original. Anything going to a funder, a client or the public deserves more care than a private brainstorm.

Always check against something outside the AI. Asking the same AI "are you sure?" is weak, because it may simply agree or change its answer to please you.

## Example

A finance assistant uses AI to summarise a 30-page audit. Before sending the summary to the board, she picks the three figures that matter most and finds each one in the original PDF. One was rounded in a misleading way; she fixes it.

## How it connects

- The answer to **hallucination**, and a guard against **sycophancy**.
- The **jagged frontier** tells you where checking matters most.
- Done systematically and repeatedly, it becomes **evals**.

## Try it

Take one AI answer you used this week and, without rereading it, write down the claim that would hurt most if it were wrong. Check that claim against a real source.