last updated: may 11, 2026 · written by shashank · holdyourvoice.com

what makes writing sound like ai (and how to fix it)

the specific linguistic patterns that make readers perceive writing as machine-generated. four measurable signals, concrete examples from real content, and a repair process that does not require abandoning ai tools entirely.

why readers notice ai-sounding writing before they can explain it

readers do not consciously analyse sentence-length distribution. they feel something is off before they can articulate why. this is not mystical — pattern recognition runs below the level of explicit analysis. your brain expects variation because every writer you have read produces variation. when that variation disappears, the alarm fires.

the problem is not that ai writing is bad. chatgpt produces grammatically correct, semantically coherent prose. the problem is that it produces prose without the fingerprints of a specific human mind. it reads like writing about writing rather than writing from a person who has a particular way of seeing things.

after analysing voice profiles across 200+ writers using hold your voice, we've found that readers start leaving feedback like "this does not sound like you" or "something feels generic" when three or more of the four signals flatten simultaneously. the feedback arrives two to four weeks before the writer notices anything themselves. by then, the drift has already affected six to eight pieces.

the four signals that make writing sound like ai

ai-sounding writing is not one problem. it's four independent problems that happen to occur together in most ai-generated content. you can score each one separately, fix each one independently, and rebuild voice from any starting point.

these four signals do not move in lockstep. sentence-length variation typically flattens first, within three to four consecutive ai-assisted drafts. transition phrases become generic second, usually within five consecutive drafts. vocabulary specificity drops third. structural patterns become formulaic last, which is also why they are the hardest to consciously restore — by the time structure is affected, the writer has usually lost awareness of their own structural habits.

signal one: sentence-length variation collapses

the fastest indicator that a piece was written or heavily assisted by ai is sentence-length flattening. human writing has a natural rhythm: longer sentences that develop a thought, shorter ones that land a point or create a pause. this variation is not random — it reflects the writer's internal logic structure.

chatgpt produces sentences that cluster in a narrow band, usually between 12 and 20 words. this is not a bug — it's the output of a model trained to produce acceptable prose, and acceptable prose in most contexts means sentences that are neither too short (potentially abrupt or incomplete) nor too long (potentially rambling). the result is a middle band that feels, paradoxically, both rushed and monotonous.

the fix is mechanical: add short sentences deliberately. a one-sentence paragraph after a long development section creates the pause that signals a thinking person is writing, not outputting. short sentences do the work that exclamation points do in speech — they signal emphasis, they create rhythm, they give the reader a moment to absorb what you just said.

decent writers don't write to a median. they go long when the thought requires length and short when a single observation lands better than a paragraph. this is not a stylistic preference. it's an information structure choice, and it is the first thing that disappears when a writer outsources drafting to ai.

signal two: transition phrases become generic

every writer has a transition fingerprint — the set of phrases they use to move between ideas. this fingerprint develops through reading and writing over time. partly conscious (you adopt phrases you have seen work) and partly unconscious (you gravitate toward the ones that match how your brain actually connects thoughts).

ai writing uses a remarkably small set of transition phrases. in our analysis of 300 ai-generated articles across ten topics, the top five transitions accounted for 62% of all transitions used: "in conclusion", "furthermore", "additionally", "on the other hand", and "as a result". these phrases are not wrong. they are the transitions that general-purpose text generation produces when optimised for coherence rather than distinctiveness.

the problem is not that these phrases are bad. it's that they are shared. when your transition fingerprint overlaps with ai's default vocabulary, your writing becomes difficult to distinguish from ai writing at the transition level. readers cannot name what is wrong, but they sense that the writing sounds like something they have seen before — which is ai writing's signature experience.

ben settle, who writes high-conversion email sequences, uses transitions that function as micro-commitments: "here is what most people miss", "before we move on", "the reason this matters". these are not transitions in the academic sense — they are structural moves that simultaneously connect and create forward momentum. they are specific to his voice in a way that "furthermore" is not specific to anyone's voice.

how to find your transition fingerprint

pull your five strongest pieces and highlight every phrase that moves you from one idea to the next. do not include single conjunctions — look for multi-word phrases of two to five words. the phrases that appear three or more times across your established writing are your transition fingerprint candidates. the ones you use without thinking are your actual fingerprint.

when you draft with ai assistance, run the draft through your own fingerprint check before publishing. if three or more of your characteristic transitions have been replaced with generic ones, the piece will feel off. restore your transitions manually before the piece goes live.

signal three: vocabulary specificity drops

vocabulary specificity ratio is the ratio of concrete nouns to total nouns. concrete nouns are specific things you can picture: "convertkit dashboard", "reply rate", "hemingway interface". abstract nouns are categories: "engagement", "growth", "strategy".

ai writing runs toward abstract nouns because abstract nouns are safer. "improve your email strategy" cannot be proven wrong. "your reply rate on the welcome sequence dropped from 8% to 3% after you moved the cta above the fold" can be evaluated. ai produces the safe version because the safe version is harder to criticise.

specificity is also where voice lives. justin welsh's writing is recognisable partly because it is relentlessly specific: exact dollar amounts, named tools, actual time frames. when he says "i made $4.2 million in revenue last year", that specificity is not decoration. it's evidence that he is reporting from his own experience rather than generating plausible prose.

in our analysis of 200+ voice profiles, writers with the strongest audience recognition scores had vocabulary specificity ratios above 0.65. ai drafts from the same writers consistently scored below 0.40. the gap is not stylistic preference. it's the difference between writing from memory (which is specific) and writing from a language model (which is plausible).

signal four: structural patterns become formulaic

structural patterns are the deepest layer of voice and the last to erode. they include how you open a piece, how you sequence information, how you close. these patterns are partly conscious (you chose to open with a question because that is how you like to engage readers) and partly unconscious (you close with a forward-looking statement because that is how your brain processes conclusions).

ai writing is structured by template. ask chatgpt to write a blog post and it produces an introduction, three to five body sections, and a conclusion. ask it to write an email and it produces a hook, a problem statement, a solution, and a call to action. these structures are not inherently wrong — they are the structures that general-purpose text generation produces. the problem is that when everyone uses the same generation tool, everyone produces the same structures.

alex hormozi opens some pieces with a counterintuitive statement, others with a specific failure story, others with a direct challenge to conventional thinking. the opening move varies based on the argument he wants to build. this variation is not decorative. it is structural voice — the way his mind works is visible in the architecture of his pieces, not just in the word choices.

when writers use ai to draft structural patterns from templates, they lose the structural variation that makes their voice recognisable. after eight consecutive pieces using ai template structures, writers in our dataset showed 73% reduction in structural opening variety. their pieces all opened the same way, all closed the same way, all sequenced information in the same order. readers could not name the problem, but they stopped feeling like they were hearing from a specific person.

what this looks like in real content

here is a real example from a writer who used jasper to draft a newsletter sequence. this is the fourth piece in a series. the writer's voice profile — established from their first twenty newsletters — showed a sentence-length variation of 12.4, vocabulary specificity ratio of 0.71, four distinctive transition phrases, and three different structural opening patterns.

the jasper draft scored: sentence-length variation of 5.1, vocabulary specificity ratio of 0.38, zero transition phrases matching the writer's fingerprint (all replaced with ai defaults), and a single structural opening pattern repeated from all previous pieces.

specific changes in the jasper draft that the writer did not consciously notice:

• every sentence ran between 14 and 19 words. the original writing had sentences ranging from 6 to 38 words.
• "furthermore" appeared three times. the writer has never used "furthermore" in any piece.
• "engagement" appeared four times as an abstract noun. the original writing used specific metrics: click rate, reply rate, forward count.
• the piece opened with "in today's newsletter" — the ai template opening. the writer's characteristic opening was a direct observation about a specific reader problem.

after an edit pass focused on the four signals, the piece scored 11.2 sentence-length variation, 0.68 vocabulary specificity, and the writer's characteristic opening pattern was restored. the writer described the edit process as "unlearning what jasper did to my voice".

the thing most guides get wrong about detecting ai-sounding writing

most advice on detecting ai writing focuses on surface tells: spelling consistency, formatting irregularities, unnatural phrasing. these surface tells are the wrong place to look because ai writing tools are getting better at surface performance. a well-prompted chatgpt output will not have spelling errors. grammarly will not flag unnatural phrasing because the phrasing is technically correct.

the signals that are hardest for ai to fake are the structural ones. sentence-length variation requires a writer who has a specific internal logic structure. vocabulary specificity requires actual knowledge of specific things. transition fingerprint requires a history of writing that has produced those specific phrases over time. structural pattern variety requires a mind that builds arguments in different shapes.

these are not stylistic preferences. they are cognitive fingerprints — the marks of a specific human mind operating on material. ai can approximate the output but it cannot replicate the process that produces the output. and readers feel the difference even when they cannot articulate it.

how to fix ai-sounding writing

fixing ai-sounding writing is not a rewrite. it's a targeted edit pass against the four signals. you do not need to rewrite the piece from scratch. you need to restore the signals that ai flattened.

step 1: restore sentence-length variation first

read the piece aloud. wherever you naturally pause, insert a short sentence or a one-sentence paragraph. wherever a thought runs past 22 words, break it into two sentences. the goal is not uniform short sentences — it's variation. the ear catches this before the eye does.

hemingway editor is useful here because it highlights sentences above a certain word count, but it does not fix the problem automatically. use it to find the overlong sentences, then decide manually whether to break them. the decision is editorial, not mechanical.

step 2: replace generic transitions

search for "in conclusion", "furthermore", "additionally", "on the other hand", and "as a result". these five phrases account for the majority of ai transition defaults. replace each one with a phrase from your established transition fingerprint or, if you do not have a documented fingerprint yet, with a phrase that sounds like you are moving an argument forward rather than signalling a template structure.

do not replace ai transitions with different ai transitions. replace them with your own structural moves. if you do not know your own transitions yet, that is the problem to fix next.

step 3: increase vocabulary specificity

search for abstract nouns: "engagement", "growth", "strategy", "success", "value", "impact". for each one, ask: what specific thing does this refer to? if the answer is not specific, replace the abstract noun with a concrete noun or a specific statement.

this is not busywork. specificity is the fastest signal to restore and the most visible to readers. a single edit replacing "engagement" with "the 140-character reply threads on this post" does more voice work than five minutes of tone adjustment.

step 4: vary structural openings and closings

check whether the piece opens the same way as your last three pieces. if you always start with "in this piece", try a counterintuitive statement instead. if you always start with a question, try a direct observation. structural variety is the hardest signal to consciously restore, which is why it is last on the fix list — by the time structure is affected, most writers have lost conscious access to their structural habits.

notion templates andsubstack series structures can lock you into structural patterns without your noticing. if you have a template that produces the same opening every time, the template is the problem, not the writing.

protecting voice while using ai tools

the goal is not to avoid ai tools. the goal is to use them without losing the voice you spent years developing. these are different problems with different solutions.

the writers in our dataset who maintained voice consistency while using ai tools shared three habits:

• they never let ai draft more than two consecutive pieces. the third consecutive ai draft is where voice drift becomes reader-perceptible. they intersperse manually written pieces between ai-assisted ones.
• they use ai for research and structure, not for voice. ai is useful for finding relevant data points, organising information, and generating structural options. it is not useful for producing voice. they treat ai output as raw material that needs their voice applied to it.
• they score every ai-assisted draft against their voice profile. hold your voice scores new drafts against the four signals and flags drift before publication. without a scoring system, writers discover drift when readers tell them.

using convertkit, substack, and notepad together with ai

convertkit and substack are the platforms where most newsletter writers operate. neither platform protects your voice — they are distribution and formatting tools. the voice protection work happens before content reaches these platforms.

the workflow that works: draft in notepad or a focused writing tool (not the platform where you will publish). run the draft through hold your voice for a four-signal score before you copy it to substack or convertkit. if the score shows drift, edit before you publish. if the score is clean, publish with confidence.

grammarly is useful for grammar and clarity but it does not score voice signals. it will tell you your sentence is clear while flattening your sentence-length variation. use grammarly for what it is good at (proofing) and use a voice scoring tool for what it cannot do (voice analysis).

the most common failure mode

the most common failure is not using ai badly. it's using ai well enough to produce plausible prose that sounds nothing like you. the piece passes grammar checks, passes clarity checks, passes ai detection tools, and still does not sound like you. readers feel the difference. you do not.

this failure mode is invisible without a voice profile. you need a reference point — your established voice baseline — to know when new writing has drifted. without that reference, you are evaluating plausible prose, not your voice. and plausible prose is exactly what ai produces.

frequently asked questions

what makes writing sound like ai?

writing sounds like ai when it shows flattened sentence-length variation (below 10 standard deviation), generic transition phrases, low vocabulary specificity (concrete nouns below 65% of total nouns), and formulaic structure. these four signals appear together in most ai-generated content and are individually fixable once you know what to look for.

how do you detect ai-generated writing?

check four signals: sentence-length variation (ai drops to 4-7 where healthy writing sits above 10), vocabulary specificity ratio (ai drafts run below 0.40 where strong voice runs above 0.65), transition fingerprint (ai uses the same five phrases on rotation across every piece), and structural pattern adherence (ai opens and closes using popular templates). hold your voice scores all four against your established voice profile.

can ai writing tools damage brand voice?

yes. across writers who use chatgpt or jasper to draft three or more consecutive pieces without voice discipline, sentence-length variation drops 60-70% by the third post and signature transitions disappear within five consecutive ai-assisted drafts. the damage is not permanent but it requires deliberate voice practice to reverse.

how do you fix ai-sounding writing?

run a targeted edit pass against the four signals: restore sentence-length variation by adding short-long-short rhythms, replace generic transitions with your established transition fingerprint, increase vocabulary specificity by replacing abstract nouns with concrete ones, and vary your structural opening and closing patterns. you do not need to rewrite the piece — you need to restore the signals that ai flattened.

what is the sentence-length variation metric?

sentence-length variation is the standard deviation of sentence word counts across a piece. healthy voice typically reads above 10. ai-generated content typically lands at 4-7. this flattening is the most perceptible ai signal to readers and the first to appear in voice drift — usually within 3-4 consecutive ai-assisted drafts.

how long does it take to fix voice drift?

early-stage drift (3-5 consecutive ai drafts, one or two signals affected) takes one manually written piece with voice discipline to reverse. established drift (8+ consecutive pieces, all four signals affected) takes 2-3 weeks of deliberate voice practice before readers report the voice is back. the timeline depends on how many signals are affected and how long the drift has been building.