The Generic Filter: Auditing a Week of AI-Drafted Social Posts

Marketing Academy · Field Work●AI in Marketing

MiniAudit· 30 minutes

Wise (formerly TransferWise)

Objective: Given a week of AI-drafted social captions from a real workflow, apply the lesson's Stage 4 human review checklist to flag brand-voice drift, hallucinated claims, and vague phrasing before anything gets scheduled.

You're the solo social media marketer at Wise, the London-founded cross-border money-transfer company (LSE: WISE), reviewing a batch of 8 AI-drafted LinkedIn and Instagram captions before Thursday's scheduling window.

Run each caption through the four-point Stage 4 checklist (brand voice, accuracy, specificity, cliche removal), decide ship, edit, or kill for each, and flag the one caption with a hallucinated statistic.

Before you start

What you'll need

Free path (everything below is enough to finish)

Google Sheets

FreeScore each caption against the 4-point checklist

Free, tabular, and easy to share with a second reviewer

The process

1 step

Step 01 of 01

Stage 4 human brand review checklist

The lesson's Stage 4 requires every AI draft to clear four checks before scheduling: brand voice, accuracy, specificity, and cliche removal.

Caption 5 claims 'transfers are now 40% faster than traditional banks, according to our 2026 customer survey.' No such survey exists in your content brief. What do you do with this caption?

Google Sheets— Paste all 8 captions into one sheet, one row each, with a column for each of the four checks.

Procedure

Import the 8 captions into rows 2-9
Score each caption pass/fail on brand voice, accuracy, specificity, and cliche removal
Isolate caption 5's unsupported statistic as an accuracy failure
Mark ship, edit, or kill for each row based on failure count

Sample output

CAPTION AUDIT (excerpt)

#5 - Instagram, 'Send money in seconds...'
  Accuracy: FAIL - cites a '2026 customer survey' not in the brief
  Verdict: KILL until the stat is sourced or removed

#2 - LinkedIn, 'International payroll...'
  Cliche: FAIL - opens with 'In today's global economy'
  Verdict: EDIT - cut the opener, keep the body

...6 more rows

Healthy

Every hallucinated stat gets caught before scheduling; only 1-2 of 8 captions ship untouched.

Unhealthy

All 8 captions marked 'ship' because they read fluently, even though one invents a source.

What this means

Fluent AI output is not the same as accurate AI output; the accuracy check exists specifically because hallucinations read confidently.

So what do I do about it?

Symptom	Action	Effort
A caption cites a stat you don't recognize	Kill or hold the post until the source is verified in your own data	5 min

YouYou can do this yourself, no engineering access required.

Final deliverable

A scored audit of 8 AI-drafted captions with ship, edit, or kill verdicts and the hallucinated stat flagged for removal.

See a reference example

Sample output

Notion, week-of Aug 18 caption audit (excerpt)

SHIP AS-IS (2)
  'Async work isn't lazy work...'

EDIT (5)
  'In today's fast-paced world, teams need...' -> cut opener

KILL (1)
  'Our Q3 report shows 340% growth' -> no such report exists, remove until sourced

Success criteria

You're done when you can:

Correctly flags the hallucinated statistic
Applies all 4 checklist dimensions to every caption
Produces a clear ship/edit/kill verdict per row