The Draft Reads Fine Until Someone Else Reads It
The short answer: AI copy sounds like AI when it names categories instead of examples and has nothing specific to say. Write the first sentence yourself, bring the specifics, and edit for rhythm at the end, and the draft reads like a person decided what it was for.
Generated copy has a strange property. It looks correct on the screen and hollow in the inbox.
The grammar is clean. The structure is sensible. Nothing is wrong in a way you could circle with a red pen. Yet readers describe the same feeling, usually as bland, corporate, or oddly weightless.
That reaction is not mysticism. It comes from a small set of specific, repeatable patterns, and each one has a fix that takes a minute or two.
This guide names five tells, explains why models produce them, and gives the prompt change and the edit that removes each. The goal is not to disguise the tool. The goal is copy worth reading.
Tell One: Every Sentence Weighs the Same
- ● Rhythm is the first giveaway
- ● Hedging replaces commitment
- ● Specifics get flattened to categories
Read a generated paragraph aloud and listen to the rhythm rather than the meaning. The sentences arrive at similar lengths, in similar shapes, carrying similar weight.
Human writing is lumpy. A long winding clause gets followed by three words. A throwaway aside sits next to the most important claim in the piece.
Models smooth that out because they optimize for the most probable next token, and the most probable sentence is a medium one. Over a paragraph the effect compounds into a flat hum.
The fix is mechanical. Find the longest sentence in each paragraph and split it, then find the most important claim and shorten it until it stings.
Tell Two: Hedging Where a Human Would Commit
Generated copy loves the safe middle. Things “can be” useful, “may help,” and “often depend on your needs.”
Some hedging is honest, especially around pricing and results nobody has verified. Most of it is not. It appears where a writer with actual knowledge would simply say which option is better and why.
The pattern comes from training on text that rewards caution, plus safety tuning that discourages confident claims. The result reads as though nobody is home.
Fix it by forcing a verdict. Ask yourself what you would tell a friend over coffee, then write that, and keep hedges only where you genuinely do not know.
Tell Three: Vocabulary Nobody Says Out Loud
Certain words cluster in machine drafts far more than in speech. Delve, leverage, robust, seamless, tapestry, landscape, unlock, and elevate all show up with unusual frequency.
The words are not banned. They are simply overrepresented, and readers who see a lot of generated content have started to notice.
More useful than a blocklist is a test. Would you say this word out loud to a customer without feeling slightly ridiculous? If not, replace it with the plain version you would actually use.
The plain version is almost always shorter, which helps with the rhythm problem too.
Tell Four: The Structure Announces Itself
Machine drafts arrive pre-organized in a way that human first drafts never are. Every section opens by restating the heading, every list has exactly three items, and every paragraph ends with a summarizing sentence.
That tidiness is a tell because real writing has uneven sections. Some points need six sentences and some need one, and a heading rarely gets restated because the reader just read it.
The most conspicuous version is the closing sentence that adds nothing. “By following these steps, you can improve your results” is a sentence that exists to end a paragraph.
Delete every summarizing line that repeats what came before it. The paragraph almost always improves, and the piece gets shorter without losing content.
Tell Five: Specifics Replaced by Categories
This is the tell that matters most, and the only one a better prompt cannot fully solve.
Ask a model about pricing and it describes tiers. Ask about a workflow and it lists stages. The answer is structurally right and specifically empty, because the model does not know your customer, your numbers, or the thing that went wrong last March.
Readers feel this immediately even when they cannot name it. It is the difference between an article that could have been written about anyone and one that was written by someone.
The fix is the only real work in this whole process. Add one detail per section that only you could supply, whether that is a customer objection you keep hearing or a step you learned the hard way.
The Tells and Their Fixes, Collected
- ● Fix the prompt before the draft
- ● Cut anything a rival could publish
- ● Add one fact only you have
The table below puts the five tells next to their cause and their fastest remedy.
| Tell | What it looks like | Why the model does it | Fastest fix |
|---|---|---|---|
| Flat rhythm | Every sentence a similar length | Probable tokens make average sentences | Split the longest, shorten the key claim |
| Reflexive hedging | “Can be,” “may help,” “depends” | Cautious training and safety tuning | State a verdict, keep only honest hedges |
| Overused vocabulary | Delve, leverage, seamless, robust | Overrepresented in training text | Say it out loud, use the plain word |
| Announced structure | Restated headings, tidy threes | Format patterns learned from web articles | Cut summarizing lines, vary section length |
| Category instead of example | Tiers, stages, general benefits | The model lacks your specific knowledge | Add one detail only you could supply |
| Missing point of view | Nothing anyone would disagree with | Averaged text has no position | Name the trade-off you actually accept |
The first four are editing problems and take minutes. The last two are knowledge problems and take thought, which is exactly why they are the ones that distinguish good copy.
If your workflow only ever addresses the first four, the writing gets smoother and stays forgettable.
Fixing It in the Prompt Instead of the Edit
Several tells can be prevented rather than repaired, and prevention is cheaper.
Name the reader precisely. “A solo bookkeeper with twelve clients who has never used automation” produces different copy from “small business owners.”
Supply a sample of your own writing and ask the model to match its rhythm and vocabulary. This works better than any instruction about tone, because the model has something concrete to imitate.
Ban the vocabulary explicitly and forbid summarizing closers. A short list of banned words in the prompt removes an entire editing pass. Our best AI writing tools guide covers which products let you save these instructions as a reusable brief.
The Editing Pass That Actually Works
- ● Read the first line out loud
- ● Break the sentence-length pattern
- ● Replace categories with examples
Three passes, in this order, handle nearly everything.
First, delete. Remove every sentence that could appear unchanged in a competitor’s article. This is brutal, and it is the single highest-value edit available on a generated draft.
Second, add. Put one specific detail into each section, from your own experience or your own records. This is where the piece stops being interchangeable.
Third, read aloud. Rhythm problems are inaudible on screen and obvious in the mouth. Fix anything that makes you run out of breath or trail off.
A grammar tool helps with the mechanics but cannot see any of this. Our Grammarly vs ChatGPT for writing comparison explains where each one belongs in the pipeline.
What Detectors Do and Do Not Prove
People reach for detection tools at this point, and it is worth being clear about what they measure.
Detectors estimate statistical likelihood, not authorship. They flag text that looks predictable, which includes plenty of careful human writing, and they can miss machine text that someone edited properly.
That makes a score a weak signal rather than evidence. Using one to accuse a writer, a student, or a contractor is a mistake with real consequences, and the tools themselves usually say so in their own documentation.
GPTZero’s FAQ is a useful example. It says that when confidence is marked high, 99.1% of human articles are classified as human and 98.4% of AI articles as AI. Even so, it warns that the results should not be used to punish students.
Use them, if at all, as a nudge toward the fixes above rather than a verdict. Our AI detectors vs plagiarism checkers guide explains why the two tools answer completely different questions.
Which Rescue Level Fits Your Kind of Writing
Internal notes and summaries: Almost none. Nobody needs rhythm in a meeting recap, and time spent polishing it is wasted. Check the facts and move on.
Customer-facing marketing copy: All three editing passes, every time. This is the writing where a generic sentence costs you a sale, and where the specific detail earns one.
Published articles competing in search: The full process plus the knowledge work. Google’s guidance warns that generating many pages without adding value for users may violate its spam policy. That makes the “add one specific detail” pass mandatory.
Emails to people who know you: Light touch, but fix the vocabulary. Nothing damages a relationship email faster than a word the recipient knows you would never say.
Technical documentation: Prioritize accuracy over rhythm. Flat, even sentences are acceptable here, but category-instead-of-example is fatal, since the whole point is the specific.
Write the Ugly First Line Yourself
One habit prevents more of this than any editing checklist. Write the first sentence of every piece yourself, before the model sees anything.
That sentence sets the rhythm, the vocabulary, and the point of view for everything the model generates after it. It also forces you to decide what you actually think, which is the input no tool can supply.
Everything in this guide reduces to one idea. The problem was never that a machine wrote the draft, and no amount of rewriting fixes a piece that has nothing specific to say.
Bring the specifics, let the tool handle the typing, and edit for rhythm at the end. That order produces copy that reads like a person because a person decided what it was for.
FAQ
What is the biggest giveaway that copy was written by AI?
The most common giveaway is rhythm. Generated sentences tend to land at a similar length and a similar level of importance, so nothing in the paragraph rises or falls. Human writing is lumpy, with short blunt lines next to long ones.
Can a better prompt fix this without editing?
Partly. A prompt that names the audience, the format, the forbidden words, and a sample of your own writing removes several tells before they appear. It cannot supply the specific detail only you know, which is the tell that matters most.
What is the fastest edit that improves an AI draft?
Delete every sentence that could appear in a competitor's article unchanged. That single pass removes most generic filler, and what survives is either specific to you or genuinely necessary.
Do AI detectors prove whether text was machine-written?
Not reliably. Detector scores are probabilistic, they flag plenty of human writing, and they can miss edited machine text. Treat a score as one weak signal rather than evidence, and never as grounds for an accusation.
Is it possible to make AI writing completely undetectable?
Yes, and that is usually the wrong goal. Aim for copy that carries information a reader cannot get elsewhere, because that survives both a skeptical reader and a search engine. Sounding human is a side effect of having something specific to say.
Sources
- GPTZero: Frequently asked questions — checked 2026-10-07
- Google Search Central: Guidance on using generative AI content on your website — checked 2026-10-07
Some links may be affiliate links. We may earn a commission at no extra cost to you.
How this article was made: it was drafted with AI assistance from the official sources listed above, and its figures were checked against those pages on the dates shown. It is not based on personal hands-on testing.
Comments
Post a Comment