AI writing sounds like AI because the model is optimising for the safest plausible next sentence, and safe prose has a signature: hedged claims, three-item lists, sentences of near-identical length, a thesis restated at the end of every section, and — the fatal one — no specifics that could only have come from someone who was actually there. The fix is not a "humanizer" tool. Humanizers paraphrase the surface of bad writing and leave the badness in place. You fix AI writing at the source: with real inputs, a stated position, banned constructions, and your own edit.
Here is the full diagnosis, tell by tell, then the fix for each.
The tells
1. Specificity is missing, and everything else follows from that
This is the root cause. Nearly every other tell is a symptom of it.
A model asked to write about, say, customer onboarding without being given any facts about your customers, your product, or your onboarding will produce prose about onboarding-in-general. It cannot name the client who churned in week two. It cannot say the thing your support lead said in Slack on Tuesday. So it reaches for the abstractions that fit every case, which is why the writing feels like it's about nothing.
You cannot prompt your way out of an empty input. Read that twice, because it is the entire article.
2. The tricolon reflex
Three-item lists, everywhere. "Fast, reliable, and scalable." "It saves time, reduces cost, and improves quality." Human writers use tricolons for rhythm, occasionally. Models use them because three parallel items are a highly-probable completion, so they arrive constantly — often with a third item that adds nothing, having been generated to fill the slot rather than to make a point.
The tell: when you can delete the third item and lose no meaning, it was never a real item.
3. "It's not just X, it's Y"
Along with "not only… but also," this is the most recognisable AI construction in English right now. It has the shape of insight — a reversal, an elevation — without any of the content. It is a rhetorical move performing profundity while asserting nothing falsifiable.
4. Uniform sentence length
Read three paragraphs of model output aloud and listen to the metronome. Sentence, sentence, sentence, all fifteen to twenty-five words, all built the same way. Real prose has a short sentence that lands. Then a longer one that carries the qualification, the aside, the evidence, and the second clause you didn't see coming. The variance is the voice.
5. Hedging fog
"Can potentially help." "May contribute to improved outcomes." "Is often considered one of the more effective approaches." The model is trained not to be wrong, so it says nothing that could be wrong. The result reads as though written by a committee protecting itself from a lawsuit.
6. Restating the thesis at every closing
Each section ends with a sentence summarising the section you just read. It is a structural verbal tic, and once you notice it you cannot stop noticing it. Human editors cut those. They are the paragraphs a reader's eye already skipped.
7. Transitional throat-clearing
"Moreover." "Furthermore." "In conclusion." "It's important to note that." "When it comes to." Almost every one of these can be deleted with zero loss — and "It's important to note that X" is nearly always better as just "X."
8. Reflexive false balance
Every claim immediately neutralised by its counterweight. "While some argue X, others contend Y." The model refuses to land anywhere, and the reader — who came for a decision — leaves with the same uncertainty they arrived with. No position, no reason to have read.
The tells, and what to do about each
| The tell | Why the model does it | The fix |
|---|---|---|
| No specifics | You gave it no facts to be specific about | Feed it real inputs: notes, transcripts, numbers, names |
| Tricolons everywhere | Three parallel items are a high-probability completion | Ban them in the prompt; delete any third item that adds nothing |
| "It's not just X, it's Y" | The shape of insight without the content | Explicit ban list in every prompt |
| Uniform sentence length | Median sentence, sampled repeatedly | Ask for deliberate variance; a 4-word sentence somewhere |
| Hedging fog | Trained to avoid being wrong | Force a position: "commit; no 'may', 'can', 'often'" |
| Thesis restated per section | Structural tic from summarisation training | Cut every closing summary sentence in the edit |
| "Moreover", "Furthermore" | Filler that satisfies coherence patterns | Delete on sight; the logic usually survives fine |
| False balance | Trained toward neutrality | Ask for the strongest version of one side, then argue it |
Swipe the table sideways to see every column →
Why "humanizer" tools are a bad bet
This is where most articles on this keyword sell you something. Here is the argument against.
They operate on the wrong layer. A humanizer takes finished text and swaps synonyms, restructures clauses, and varies sentence length. That is surface texture. The reason your draft is dull is that it contains no information a reader couldn't get elsewhere — and no amount of clause-shuffling adds information. You end up with prose that is differently mediocre, and frequently with odd word choices where a thesaurus swap replaced a precise term with an approximate one.
They optimise against a broken referee. The premise of most of these tools is beating AI detectors. But AI detectors are unreliable in both directions — they flag human writing as machine-written, and they miss machine writing, and their behaviour changes without notice as models change. Optimising your prose against a noisy, moving target is not a strategy. It is a way to spend money making your writing worse in a direction nobody asked for.
They solve a problem you shouldn't have. If the goal is to sneak past a detector, the underlying question is whether the text should have been written that way at all. If the goal is good writing, detector scores are irrelevant to it. Those two goals almost never point the same direction, and if you find yourself paying a monthly fee to hide how something was made rather than to make it better, that is worth sitting with.
If you take one thing from this piece: no tool fixes bad writing after the fact. The leverage is entirely upstream.
Fixing it at the source
Give it something only you have
The single highest-leverage change. Before you prompt, gather:
- Your own rough notes, however messy — including the half-formed opinion you'd be slightly embarrassed to publish
- Real transcripts: a support call, a sales call, a customer interview
- Actual numbers from your own systems, with their context
- The specific counter-argument someone made to you, and who made it
Then prompt with the material, not the topic. "Write about churn" gets you a Wikipedia article. "Here are twelve support tickets and my notes from the exit interviews; find the pattern and argue for it" gets you something nobody else could have written. The difference in output is not incremental.
Make it commit
Add this to your prompt, more or less verbatim:
Take a clear position and defend it. Do not present both sides. Do not hedge — no "may", "can", "often", "generally". If you're uncertain, say so once, explicitly, and then commit anyway.
The refusal to land anywhere is the deadest quality in machine prose, and it is directly promptable away. Most people never try.
Ban the constructions
Paste a ban list into every writing prompt. Mine:
Banned: three-item lists; "it's not just X, it's Y"; "not only… but also"; "moreover"; "furthermore"; "in conclusion"; "it's important to note"; any clichéd scene-setting opener about the modern world or the evolving landscape; any sentence that summarises the section it ends. Vary sentence length deliberately — include at least one sentence under six words per section.
Models follow negative constraints imperfectly, so you will still catch a few in the edit. But it removes most of them, and it removes exactly the ones a reader clocks first.
Give it your voice, not a description of your voice
"Write in a friendly, professional tone" is close to useless — it describes a region of style-space so large as to be meaningless. Paste in 800 words you actually wrote and say "match this: the rhythm, the sentence length variance, the vocabulary, the willingness to be blunt."
This is where the brand-voice tools have a real, narrow use. Jasper and Writesonic both maintain a voice profile across many outputs, which is genuinely valuable when a team is producing volume and consistency is the goal. What they will not do is make dull content interesting — nothing does, except having something to say. For long-form reasoning and argument, a strong general model like Claude with real inputs and a ban list will outperform a dedicated marketing generator most of the time. The rest of the field is in AI writing tools and AI copywriting tools if you want to compare, and our roundup of the best AI writing tools covers who each one is genuinely for.
Edit by cutting
The last pass, and the one everybody skips:
- Delete the first sentence of every section. It is throat-clearing eight times out of ten.
- Delete every closing summary sentence.
- Find every tricolon. Cut the weakest item. If cutting it loses nothing, cut a second.
- Search for "may", "can", "often", "generally", "potentially". Delete or commit.
- Read one paragraph aloud. If it sounds like a metronome, break a sentence in half.
- Add one specific thing per section that could only have come from you — a name, a number, a moment, a thing someone said.
Step six is the whole ballgame. Everything else is hygiene.
FAQ
What AI writing tools don't sound like AI?
The tool matters far less than the input. A strong general model such as Claude or ChatGPT, given real source material, a clear position to argue, and a ban list of AI constructions, will out-write a dedicated content generator running on a topic keyword alone. Purpose-built tools like Jasper and Writesonic are strongest at holding brand voice steady across volume, not at making a thin idea worth reading.
Do AI humanizer tools actually work?
They change the surface of the text — sentence structure, word choice — without adding anything a reader wants. They also frequently introduce imprecise vocabulary in the swap. And since they are usually sold on the promise of evading AI detectors, which are unreliable in both directions and change constantly, you're paying to optimise against a moving and inaccurate target. Fix the draft instead.
Can AI detectors be trusted?
Not enough to make decisions with. They produce false positives on human writing — particularly writing that is plain, structured, or by non-native English speakers — and false negatives on machine writing. Treat a detector score as weak evidence at best, and never as the basis for an accusation or an editorial standard.
How do I make AI writing sound more like me?
Give it samples rather than adjectives. Paste in several hundred words you genuinely wrote and ask it to match the rhythm, sentence-length variance, and vocabulary — not to "be professional and friendly," which means nothing. Then edit by cutting, and add at least one specific detail per section that only you could know.
Is it obvious when content is written by AI?
To a careful reader, usually — but they rarely notice the machine. They notice that it says nothing. Content gets clocked as AI when it has no specifics, no position, and no risk in it, and a human writing carelessly produces exactly the same signature. Specificity is what reads as authenticity.
Write something only you could write
The uncomfortable conclusion is that most AI writing sounds like AI because most of it was produced without anything to say. The model is not the problem. It faithfully reflects the emptiness of the input.
So before reaching for another tool: gather the real material, take a position, and cut the writing that is protecting you rather than informing anyone. If you want to compare what's out there once your process is right, AI writing tools and AI copywriting tools are the places to start, and the guide to AI tools for content creators covers the rest of the stack. If you're taking one strong article and stretching it across channels, the repurposing workflow has the prompts for it. Building a tool that helps with the diagnosis rather than the disguise? Submit it.