Detection tests
Run every one that applies — all seven for a team running several accounts, the first six for a single author publishing alone. Each catches a failure the others miss. A draft that fails any test goes back through the pass, and the pass repeats until the battery comes back clean.
1. The squint test
Look at the shape without reading: paragraph lengths, line breaks, bold density, list frequency.
Machine drafts have machine silhouettes. If the shape alone suggests AI, readers decided before word one.
Cheapest test in the battery and the one most worth running first, because it catches the staircase and uniform rhythm in about three seconds.
2. The read-aloud test
Machine rhythm surfaces immediately when spoken: identical sentence lengths, stacking parallels, clauses no speaking human would chain.
Anywhere breath runs out or the cadence turns metronomic, rewrite.
This is also the fastest way to catch register drift, since a feed post read aloud in an essay register is obvious to the ear and invisible to the eye.
3. The skeleton x-ray
Replace every noun and specific with a blank and look at what remains.
If the skeleton alone identifies the machine, flip, triad, triad, summary close, then the surface edits never mattered.
This is the test that catches the failure the whole method is built around: a draft where every suspicious word was swapped and the structure was left intact.
4. The blind lineup
Place the draft next to three genuine pieces by the same author and ask someone who reads that author weekly to find the fake.
Chance-level performance is the pass bar.
This is the definitive voice test. It is also the most expensive, so run it on the pieces that matter and on a sample of the rest. If it is only run once per author, run it early, because a failure here usually means the fingerprint is missing rather than the draft is bad.
5. The competitor test
Could a direct rival publish this piece unchanged, logo swapped?
If yes, it argues for the category and says nothing about the company. Inject the specifics only this operator has, or do not ship.
Most drafts that pass every other test fail this one, because avoiding tells is a subtractive skill and differentiation is an additive one.
6. The cost test
Find one line that costs the author something: a stance a peer could challenge, an admission, a number they will have to defend.
If no line costs anything, the piece is upholstery. Machines produce costless content by design, because nothing is at stake for them.
The line does not need to be dramatic. A named tradeoff, a decision the author regrets, or a number that invites scrutiny all qualify.
7. The correlation check
For teams running several accounts: chart each account's reach baseline.
Synchronized rise and fall means a shared production signature the feed has priced in, and no single post's quality will fix it. Break the shared element, then re-baseline.
Skip this one when a single author is publishing alone. Run it monthly otherwise, because it is the only test in the battery that operates above the level of a single draft.
Repeated failures
When a draft keeps failing the lineup or the x-ray after repeated passes, stop editing.
The problem is upstream. Either the source material is too thin to support specifics, or the voice fingerprint is missing and every pass is guessing at the author. No amount of polish substitutes for either.
Go collect source material or build the fingerprint, then start the pass again from step one.