SEO & Growth Hacking with AI
0/24 complete

Module 04 · Content Production at Scale

Editorial QA: Catching AI Hallucinations Before Publishing

20 minfocused lesson5practical steps4grounded questions3source links
Open lesson + course map

On this lesson

Course outline

Editorial QA must prove that each factual claim is supported, current enough, accurately represented, and safe for the audience. Fluent language is not evidence. Review claims at sentence level and verify code/math/process independently.

// concept

Create a Claim Register

// prompt — copy me8 lines
claim_id
exact claim
type: fact / observation / sample / recommendation
source URL/document/version/date
supporting passage or test evidence
risk: normal / unstable / high-stakes
reviewer and result
update trigger

High-risk areas include prices, laws, tax, medical/health, finance, security, product limits, dates, statistics, and claims about people/companies. Use qualified review when appropriate.

// concept

Run the QA Stack

  1. Source validity: primary, reachable, relevant, correctly scoped.
  2. Entailment: source actually supports claim.
  3. Freshness: unstable facts checked near publication.
  4. Originality/copyright: no copied structure/passages or excessive quotation.
  5. Math/code: execute samples and recompute values.
  6. Safety/privacy: no private data, unsafe instruction, or hidden credentials.
  7. Commercial truth: real availability, price, location, terms, testimonial, and results.
  8. Language parity: translations preserve meaning.

AI can flag possible issues but cannot be the sole verifier of its own draft.

// worked_example

Worked Example

A draft says Google guarantees indexing if you submit a sitemap. The claim register links Google’s documentation, which contradicts it. The correction says a sitemap helps discovery and monitoring but does not guarantee crawling/indexing.

Another paragraph gives a sample 20% conversion rate without label. It is changed to a clearly marked worked example with calculation, not a benchmark. A quoted sentence is paraphrased and cited within copyright limits.

// failure_cases

Failure Cases to Diagnose

6 cases to diagnose

  • Source exists, so claim passes

    test direct support.

  • Secondary blog for current product rule

    use official docs.

  • AI reviewer agrees with AI writer

    require human/test evidence.

  • Broken source ignored

    replace or remove claim.

  • Sample number looks like market data

    label prominently.

  • Correction after publish has no log

    preserve correction state/date.

// pakistan_angle

Pakistan Angle

Use FBR, SBP, SECP, regulator, university/board, provincial authority, or provider sources where relevant. Do not generalize one province/city rule to Pakistan.

Check lakh/crore, PKR minor units, dates, phone formats, marla conventions, and Roman Urdu meaning. These details often create costly hallucinations despite fluent copy.

Maintain an update trigger for each unstable claim. A pricing statement may require a monthly check; a product feature may need review after release notes; a legal or regulator reference needs review by a qualified owner. A citation is not permanent proof when its underlying fact changes.

// hands_on

Hands-On Exercise

5 steps

  1. Extract every factual claim from one draft.

  2. build the register.

  3. verify source support/freshness.

  4. execute code/math and label samples.

  5. approve, revise, or remove each claim.

// completion_rubric

Completion Rubric

6 checks — tick as you verify

0/6

// sources

Sources

// check_yourself

Check yourself

4 questions · answers and options are taken word-for-word from this course

0/4
  1. 1 / 4 · diagnose

    Your work shows this failure mode: “AI reviewer agrees with AI writer.” What does the lesson tell you to do about it?