AI writing that doesn't sound like AI

A reader knows in the first three sentences. They cannot always articulate it. The page just has that smell. Em dashes everywhere. A "delve into" early. A three-adjective stack in the second paragraph. The little wrap-up sentence that explains what the previous paragraph meant. Once a reader catches the smell they stop trusting the writer, even if the prose is technically clean. Humanizing AI output is not about adding personality. It is about removing the small set of moves that telegraph the source.

Here are the tells, then the kit that strips them. Both halves matter. Knowing the tells without a process to remove them is just self-loathing.

The tells, listed by frequency

These show up in unedited AI output in this rough order:

• The em dash habit. AI prose loves the em dash. It uses it for asides, for emphasis, for breath, for parenthesis. A human writer uses one or two per chapter. AI uses three per paragraph. The fix is total elimination. Replace every em dash with a period and a new sentence, or a comma, or restructure the clause out entirely.
• The filler verbs. "Delve into." "Dive into." "Unleash." "Harness." "Navigate the complexities of." These are not writing. They are placeholders for thought. A human writes "Mara opened the box," not "Mara dove into the contents of the box."
• The tricolons. Three parallel adjectives in a row. "Bold, sweeping, ambitious." "Quiet, lyrical, devastating." Once is fine. Three times in a chapter is a tell. The fix is to pick one of the three and cut the others. The one you keep should be doing real work.
• The hedge openers. "It is worth noting that." "One might consider." "It should be mentioned." None of these earn their place in fiction. None of them earn it in functional prose either. Cut.
• The moralising wrap. Every paragraph or scene closes with a sentence explaining what the previous content meant. "Mara had finally understood what her grandmother meant about love." This is the model defaulting to a school-essay structure. Real fiction trusts the reader to understand.
• The "in today's fast-paced world" intros. Even when the article is not about today's anything. The model reaches for these openers because its training rewarded them.
• The hollow superlative. "Best-in-class." "Cutting-edge." "Game-changing." "Powerful." Words that mean nothing unless something specific is attached. A human writes "the fastest planner I have ever shipped." A model writes "a powerful planning tool."
• The rhythm flatline. Sentence lengths cluster. Most clauses are 12 to 18 words. The variation that gives prose its breathing comes from short fragments next to long sentences, and AI default output has neither. Fragments where they earn their place. Long sentences that actually need their length. Both, in alternation.
• The corporate cheerfulness. A tone of mild upbeat competence that hangs over everything. No real anger, no real grief, no real strangeness. The model has been trained out of intensity. The fix is to put intensity back in, by hand.

What to ban

A practical ban list, used as a final pass on any AI-assisted prose. These are the words and structures to grep for and remove:

• All em dashes.
• All instances of: delve into, dive into, harness, unleash, navigate the complexities, in today's fast-paced world, it is worth noting that, one might consider, elevate, empower, seamless, robust (as praise), game-changer, cutting-edge, leverage (as verb), supercharge, ecosystem (as fluff), comprehensive (as filler), powerful (as filler).
• Any three-adjective parallel construction. Search for ", " followed by an adjective, followed by ", " followed by an adjective, followed by " and " or ", and ".
• Any paragraph whose final sentence begins with "In essence," "Ultimately," "In the end," or "What this means is."
• Any sentence that ends with the moral of the preceding sentences.

This is not poetry. It is a search-and-destroy pass. A grep over the manuscript, a manual review of every hit, and a decision: rewrite, cut, or rare exception. The exception case should be rare enough that you can defend it in one sentence.

What to keep

The fix is not to make AI output sound like a stylised version of something else. The fix is to remove the tells and let the underlying scene stand. Some defaults from AI are fine. The model can pace dialogue tags reasonably. It can stage blocking. It can hold a POV. None of these are tells.

Do not over-correct. A piece of writing scrubbed of every common phrase becomes its own kind of artefact, equally identifiable as AI-assisted by virtue of how strenuously it avoids cliche. The goal is invisible, not weird.

The fix kit

Four moves, applied together, get most prose past the smell test.

Feed the model your own prose

The single highest-leverage move. A long sample of your finished writing, attached to every drafting call, drags the output toward your sentence rhythm and vocabulary. This is not magic. The model is doing pattern completion against the context. If the context contains your voice, the completion gets closer to your voice.

A thousand to two thousand words is the right size. More is wasted. Less is too thin to influence anything. Use a finished scene, not a draft. Use a scene that is recognisably your style at its most yourself.

Name your influences

Tell the model who you read. Not as flattery. As coordinates. "Spare, observational, in the tradition of Marilynne Robinson and Kent Haruf" is real instruction. The model has read those writers. It can target their cadence.

Also tell it what you are not. Negative constraints sharpen positive ones. "Not maximalist. No baroque clauses. No long parenthetical asides."

Ban specific phrases

Hand the model the ban list. Tell it never to write any of those phrases. The model will mostly comply. The compliance is imperfect, which is why the final-pass grep matters. But the upfront ban catches maybe 80% of the hits, and the grep catches the rest.

Run a final-pass humanize

This is where Pendraic's humanize pass lives in the pipeline. The drafting layer produces prose. A separate pass scans for the ban list, flags every hit, and either rewrites or surfaces for the writer's review. The pass also checks rhythm: clusters of similar sentence lengths, repeated paragraph closers, the moralising wrap.

You can do this manually if you are working outside Pendraic. The grep is straightforward. The rewriting takes time. If you are doing volume, automation is the difference between this step happening and not happening.

The honest moment

This is hard. Removing the tells is not a one-pass fix. It is a discipline you apply across the manuscript, and the discipline only sticks if the writer is in the loop.

The writer has to read every page out loud. The writer has to make the calls on which exception to accept. The writer has to notice when their own prose has started absorbing AI defaults, which it will, because that is how writing brains work. The tool helps. The tool does not replace the ear.

A workflow that hides this from you is not solving the problem. It is delaying the moment a reader notices.

When the tell is not a tell

A short defence of one specific case. There are writers whose voices include some of these features. The deliberate em dash user. The writer who reaches for "elevate" because she means it. The literary stylist whose tricolons are load-bearing because she is in conversation with a tradition that uses them.

If that is you, keep them. The ban list is for AI defaults, not for human choices. The distinction is whether you would have written the phrase without the model nudging you toward it. If you would, keep it. If the model put it there, cut it.

This is a real distinction and an annoying one. The fix is to know your own voice well enough to hear the difference, which goes back to the prose-sample move at the top of the kit.

Where Pendraic comes in

Pendraic builds the humanize pass into the drafting pipeline. Every customer-facing string the platform produces, and every prose call run through the editorial chain, goes through a final scan that flags AI-tell phrases, em dash habits, hollow superlatives, and rhythm clusters. The pass either rewrites or surfaces for review. The ban list ships with the platform and you can extend it per project.

If you have a draft that came out of an AI tool and you can smell the tells but cannot find them all manually, sign in and run it through the humanize layer. The first pass tends to surface more hits than writers expect.

A novel that does not sound like AI is not magic. It is the same prose with the tells removed.