A strange ritual has taken hold in content teams. A draft arrives. Someone pastes it into an AI detector. The detector returns a percentage. If the number is high, the draft goes through a “humanizer,” which rewrites it until the number drops, and then it goes back into the detector to make sure. The draft that finally gets published has been shuffled by two machines and verified by neither.
Call it detector roulette. It’s popular because it feels like diligence. It produces a number, and numbers feel like evidence. But it answers the wrong question.
The question that matters isn’t “does this text look machine-written?” It’s “is this text true, sourced, on-brief and good?” Those are editorial questions, and they have editorial answers.
Why is “was AI involved?” the wrong question?
Because it isn’t the problem readers or clients actually experience. A reader harmed by a piece of content is harmed by a false claim, a missing caveat in a sensitive topic, an invented source or a confusing recommendation. None of those harms depends on whether a model, a person or a combination wrote the sentence.
There’s a policy dimension, of course. Some clients ban AI use entirely, and that’s their right. But even then, the answer isn’t a detector score. It’s an honest process: if the client bans AI, the work is done and checked by hand, and the draft’s origin is known rather than guessed.
What have authenticity investigations historically focused on?
When it really mattered whether a document was what it claimed to be, investigators looked past surface style to substance and provenance.
The Federalist Papers. The authorship of several essays was long disputed between Alexander Hamilton and James Madison. In the 1960s, statisticians Frederick Mosteller and David Wallace published a careful study of word-use patterns that attributed the disputed papers to Madison. It was rigorous, slow and specific, with a known set of candidate authors, a long way from a percentage in a browser tab.
The Hitler Diaries. In 1983, the German magazine Stern published what it claimed were Hitler’s diaries. They were quickly exposed as forgeries, with forensic analysis of the materials, including paper and binding, playing a central role. Physical provenance settled what impressions couldn’t.
Piltdown Man. Presented in 1912 as an important human ancestor, the find was shown in 1953 to be a forgery, combining a human skull with an ape’s jaw. Testing of the materials revealed what decades of interpretation had missed.
The lesson for editors: when authenticity matters, verify the substance and the provenance. Don’t grade the vibe.
What should editors check instead?
Four things, in order of how much harm they prevent.
1. Claims
List every factual claim in the draft. For each one, record the source, the source’s date and a status: Verified, Needs source, Cut or Escalate. AI-assisted drafts tend to contain fluent, specific-sounding claims that have no source at all. A claim log finds them systematically rather than by luck.
2. Citations
Open every cited source. A source you can’t open, a study that doesn’t appear where it’s said to appear or a journal name that seems slightly off is a red flag regardless of who wrote the sentence. Invented citations are among the most damaging failures in published content, and they’re entirely checkable.
3. Risk
Score the draft’s risk on a few clear factors: unverified claims, topic sensitivity, citation quality, client AI policy and brand or legal exposure. A high score sends the draft to Hold or Escalate. A complete, low score makes it eligible to publish.
4. Editorial quality
Voice, structure, CTA and on-page basics. This is where AI-assisted drafts often need the most human work, not because they’re poorly written but because they’re generically written. The fix is editing toward the brand’s specific voice, specific examples and specific opinions.
Why are “humanizers” worse than useless for editors?
Because they change the surface while leaving the substance untouched, or make it worse. A humanizer can’t verify a claim. It can’t open a citation. It doesn’t know the client’s product. What it does is rephrase, and rephrasing can blur a precise claim into a vague one, introduce new errors or strip out the very structure that made the piece useful.
There’s also an integrity problem. Using one model to disguise another model’s output for a client who prohibited AI isn’t editing. It’s concealment. A good editorial process makes the draft’s origin irrelevant to its quality and transparent where the client’s policy requires it.
What does a substance-first edit look like in practice?
Here’s a sequence an editor can run on any draft, AI-assisted or not:
- Read once for comprehension. Don’t edit yet. Note what the piece claims to do.
- Build the claim log. Every factual claim, with a source and status.
- Open every citation. Mark anything you can’t open as Needs source.
- Score the risk rubric. Leave nothing blank. An unscored factor means Hold.
- Run the editorial pass. Voice, structure, CTA, on-page basics.
- Decide: Publish, Hold or Escalate. Record the verdict and the reason.
Notice that nothing in this list asks who wrote the draft. The process holds AI-assisted drafts and human drafts to the same standard, which is precisely the point.
How do you make AI-assisted drafts sound less generic?
Editing toward specificity does more than any humanizer. Six moves that work:
- Replace generic examples with the client’s real ones.
- Cut throat-clearing openings. Start where the useful part starts.
- Add an opinion the brand actually holds. Generic drafts avoid positions.
- Swap abstract nouns for concrete verbs. “Optimization of workflows” becomes “faster handoffs.”
- Vary sentence rhythm deliberately. Read a paragraph aloud and fix the monotony.
- Delete the summary paragraph that restates the intro. It’s almost always there.
These edits make a draft better for readers. That’s the goal. Any change in how a detector scores it is beside the point.
What about clients who ban AI?
Respect the policy and make compliance verifiable through process, not through detectors. If a client bans AI, the work is drafted and checked by hand, the editorial checklist is run on a human edit and the draft’s origin is documented. Client AI policy is one of the risk factors you score: if a client bans AI and a draft’s origin is unknown, that’s a 2, and the draft is held or escalated until the origin is resolved.
That’s an honest answer to an honest policy. A detector percentage isn’t.
What should a team’s AI-assisted drafting policy say?
Short policies get followed. A practical one fits on a single page and answers six questions:
- When is AI assistance allowed, and for which kinds of content?
- Which clients or topics prohibit it, and how is that recorded on the brief?
- What must every draft include before editing starts, such as the brief and the fact sheet it was written from?
- Who verifies claims, and where is the claim log kept?
- Which topics always escalate, regardless of how the draft was produced?
- What is never acceptable, such as invented sources, invented reviews and disguising AI output for a client who banned it?
Notice that the policy says nothing about detector thresholds. It’s built on process and records, which is what lets an editor stand behind a piece when someone asks how it was made and checked. It also gives new writers and freelancers a clear standard from day one, instead of a vague sense that “we’re careful about AI.”
How does the Draft-to-Publish Editorial Kit support this?
The Draft-to-Publish Editorial Kit is built around substance-first editing. It includes a Claim-Source-Date Log, a five-factor Risk Rubric whose formulas refuse to treat blank scores as zero, an Editorial Pass for voice, structure, CTA and SEO basics, and a Publish Decision sheet that records Publish, Hold or Escalate. A seven-page guide explains the decision rules, and a fictional before-and-after walkthrough, Harbor & Hearth, shows the whole process with annotations.
The kit takes a clear position: it’s not for evading detectors. It’s for making drafts true, sourced and good, whether a client allows AI with human verification or bans it outright.
Quick answers
Should editors use AI detectors at all?
Treat any detector result as a weak signal at most, never a verdict. Your editorial decision should rest on claims, sources, risk and quality.
Is AI-assisted content worse for SEO?
Search engines have said they focus on content quality and helpfulness rather than how content is produced. That puts the emphasis exactly where substance-first editing does.
How long does a substance-first edit take?
Longer than pasting into a detector, shorter than repairing a published error. For a typical blog post, plan on thirty to sixty minutes, depending on the number of claims.
Edit the text, not the score
Detector roulette is attractive because it’s fast and produces a number. But the number measures the wrong thing, and the humanizer that follows can make the draft worse. Editors earn their keep by checking what readers rely on: facts, sources, risk and quality.
Get the Draft-to-Publish Editorial Kit and run your next AI-assisted draft through a claim log, a risk rubric and an editorial pass instead. You’ll publish with a record of what you verified, which is worth more than any percentage.
You’ll find the Editorial Kit on Content & SEO Workflows, built for substance-first edits.
Leave a Reply