How to Spot AI Writing Tells and Edit Them Out

We all see so much generated content now that we can spot it almost immediately. The em dash where a comma would do, the "delve," the tidy three-item list, the sentence that hedges before it commits and that completely unnecessary re-statement.

Recognizing these patterns is the difference between copy that reads as yours and copy that reads as a machine's average of everyone else's.

But, we know first hand that it is possible to create content that is genuinely useful, well researched and informative with AI tooling. You just need to understand the way that the systems work, and then combat their worst instincts.

What are AI writing tells?

AI writing tells are the stylistic and content-level patterns that mark text as machine-generated. They cluster in five places:

Vocabulary — specific overused words
Punctuation — em dashes, colons, enormous amounts of semicolons
Structure — uniform sentence rhythm, formulaic openers
Tone — hedging, forced enthusiasm
Voice — the absence of a consistent perspective

None of these is proof on its own because a human can write an AI-sounding sentence by accident. I mean, AI was trained on human writing, after all!

But they can pile up, and once you know what to look for, an editor can infer the origin pretty quickly.

The practical point for anyone running a content operation is that these tells are editable. An editor can strip most of them in a pass. The deepest one, voice, takes more work, and it's the one that decides whether your brand sounds like itself or like every other company that fed the same prompt into the same model.

Why AI writing is so recognizable

AI text is recognizable because models generate it through next-token prediction. A large language model produces text one token at a time, computing a probability distribution over its entire vocabulary at each step and selecting from it. Because pre-training minimizes next-token prediction error across billions of documents, output drifts toward a statistical mean of aggregated training text rather than an individual author's perspective.

Two mechanisms are at the center of why this actually happens. This isn't crazy important for you to fully understand if you're just trying to make better content, but it's useful context.

Decoding strategy. Greedy decoding, which picks the single most probable token every time, produces repetitive, degenerate output, a failure Holtzman et al. formally analyzed in "The Curious Case" of Neural Text Degeneration. Sampling methods add variety, but the base tendency toward the predictable remains.
Alignment tuning. The InstructGPT pipeline trains a reward model on human preferences, then fine-tunes the language model to maximize that reward. Human raters show a typicality bias toward familiar, fluent, predictable text, and that preference produces mode collapse, which researchers describe as "an excessive and harmful reduction in output diversity." LLM-generated story continuations have two to four times lower entropy than human-authored fiction, and the gap widened after RLHF alignment.

These systems manifest their biases in pretty specific ways.

The vocabulary tells

Certain words appear at rates that would be statistically absurd in human writing. The clearest evidence comes from the peer-reviewed Kobak et al. study in Science Advances, which measured how much more often specific words appeared in 2024 PubMed abstracts relative to a pre-LLM baseline: "delves" at 28 times the prior rate and "underscores" at 13.8 times. A separate arXiv analysis of the same corpus tracked "delves" from 0.21 occurrences per million in 2020 to 14.38 in 2024, an increase of roughly 6,700%.

The phrase "meticulously researched" climbed roughly 3,900% in the generative-AI era.

Here is a working checklist for an editing pass. If you see these examples clustered, treat the draft as suspect:

Marker words: delve, delves, underscore, underscores, showcasing, intricate, meticulous, boasts.
Inflated adjectives: crucial, comprehensive, pivotal, seamless, robust, holistic.
Filler verbs: leverage, utilize, foster, facilitate, harness, streamline.
Signal phrases: "meticulously researched," "it's important to note," "in today's world."

One caveat is that some of these frequencies have started to decline as writers and tools adapt. A 2025 arXiv preprint on the coevolution between human writers and LLMs found that any fixed word list has a declining shelf life. So you have to move further into the matrix here.

Punctuation and formatting tells

The em dash is the most quantified punctuation tell. We're pretty pained by this, to be honest, as huge fans of the em dash. It's great for conjoining related ideas in a sentence! But, AI love loves it.

Human writers use it about 3.23 times per 1,000 words. A 2026 arXiv preprint measured GPT-4.1 at 10.62 per 1,000, more than three times the human rate, and the pattern persisted even when researchers suppressed overt markdown. G

PT-4o uses em dashes at roughly ten times the rate of GPT-3.5. In ecology journal abstracts, the relative frequency of em dashes more than doubled between 2021 and 2025, with no other character coming close to that magnitude of change.

Models vary, which matters for detection. A direct comparison found ChatGPT, Copilot, and Deepseek all made heavy use of em dashes, while Claude used only two in the same test and both Gemini and Meta.ai used none. Llama models register at 0.0 per 1,000 words in the arXiv data. So a wall of em dashes points to specific models, not to "AI" as a category.

Colons and semicolons show a similar skew. The prose-check maintainers report Claude-model colon use at 4.12 per 1,000 characters against a human rate of 1.01, a 4.1x gap, with semicolons at 3.1x. Treat that as indicative rather than definitive. The data comes from a GitHub documentation file. In practice, editors see this in colon-heavy titles ("Content Marketing: A Comprehensive Guide") and prose that reaches for a colon where a period would read cleaner.

Add the over-formatted layout to the list. If you see suspiciously symmetrical bullet lists, especially with every item the same length and bold labels applied with machine regularity.

Researchers frame the leading theory for the punctuation habit as markdown leakage, structural patterns from markdown-heavy training data bleeding into prose even when researchers strip the markdown itself.

Then, there's the rythm issue.

Sentence rhythm and structural symmetry

You can spot AI by rhythm faster than by any single word. Human writing is uneven. Some sentences surprise, others plod, and the variation itself carries a signature. AI writing is consistently smooth. GPTZero defines this property as burstiness, "a measure of how much writing patterns and text perplexities vary over the entire document." Human writing runs high on burstiness. AI writing runs low.

Models formulaically apply the same rule to choose the next word, which flattens sentence-to-sentence variation. Stylometric studies confirm LLM output shows more consistent sentence lengths and greater grammatical standardization than human text.

Consider the before-and-after. The AI version:

Our platform helps teams work faster. It streamlines their workflows. It improves their output. It empowers them to achieve more.

Four sentences, nearly identical in length, each built on the same subject-verb-object frame. Now the human edit:

Our platform kills the busywork. Teams ship faster and achieve more because it improves output by streamlining their workflows.

The second version varies sentence length deliberately and opens with a short clause against a longer one. That contrast, short against long, punchy against winding, is the texture AI struggles to produce and the first thing to reintroduce when you edit.

Structural crutches

Beyond rhythm, AI leans on a small set of structural moves that recur across drafts. Learn the four and you'll catch most of them:

Empty rhetorical openers: A section that starts with "Have you ever wondered why...?" or "What makes great content great?" The question adds nothing and stalls the point.
Transition crutches: Moreover, furthermore, consequently, and additionally strung between paragraphs to simulate logical flow where the argument doesn't earn it.
The "not X, but Y" formula: phrasing like "not just a tool but a system," where the sentence defines something by what it refuses to be. Researchers identify negative parallelism and the "it is not just X, it's also Y" construction as AI favorites.
List preambles: "There are three ways to..." or "Here are five reasons why..." announcing structure instead of delivering the first point.

Each of these substitutes a template for a decision. A human writer makes the transition because the logic demands it where a model reaches for "furthermore" because "furthermore" is what tends to come next statistically.

Strip the crutch and you usually find the sentence underneath is stronger for standing on its own. Which, it turns out, is good advice for human writers as well as agentic ones.

Tonal tells

The AI voice has a temperature, and it's always the same. It's typically polite, predictable, inoffensive and upbeat. Once again, this is caused by training methodologies.

The first is hedging. One RLHF study found human annotators assigned weakeners an average reward score of −1.86 while favoring plain statements and strengtheners, and that RLHF-tuned models emit more strengtheners than weakeners, reversing the pattern in base models.

Reward models also favor high-confidence responses regardless of correctness. The output reads as confidently vague, with sentences that sound assured while committing to nothing.

The second is sycophancy. Analysis of the hh-rlhf dataset shows matching a user's views is one of the most predictive features of human preference judgments. The presence of a single sycophantic feature shifts preference probability by up to about 6%, and roughly 30 to 40% of prompts carry a "positive reward tilt" that rewards agreement.

In copy, this surfaces as forced enthusiasm and performative agreeableness, the exclamation points and "great question!" energy that flattens brand voice into generic corporate cheer. We wish this was only an AI problem, but business writers may recognize this framing from style guides they've worked with.

Voice and perspective tells

The deepest tell is the absence of a voice, and it's the hardest to edit out. AI copy rarely commits to a first-person perspective, rarely offers a personal anecdote, and rarely holds a consistent point of view across a piece. Sam Kriss argued that even after you clean up the vocabulary, "the shapes of sentences" and paragraphs are wrong because no one is standing behind the words.

Pre-training pushes output toward a statistical mean rather than an individual author's perspective, and alignment tuning sharpens that mean into a single agreeable register. A brand voice is the opposite of a mean. It's a specific stance, a set of things this company believes and refuses to say, a way of reacting to facts that belongs to it and no one else. The model has no stance to draw from, so it produces prose that's clean, competent, and anonymous.

This is why voice is the tell that matters most for brand content.

Vocabulary and punctuation are surface features a good editor fixes in one pass where voice is structural. Restoring it means reintroducing a point of view with an opinion stated plainly, a specific example only this company would use or a consistent first-person perspective that reacts to the material instead of summarizing it.

For a content operation running at volume, this is the hardest thing to keep consistent.

Content-level tells beyond style

Hallucinated factoids are the most consequential tell beyond the stylistic stuff. Documented hallucination rates vary by model, but the range is wide enough to demand a fact-check pass on every draft. On the PersonQA benchmark, o3 hallucinated on 33% of prompts and o4-mini on 48%. Reasoning models, which make more distinct claims overall, tend to hallucinate more on recall tasks.

Flawless grammar is also a big one because human writers make spelling and grammar errors or show out-of-band preferences and models mostly don't. Grammar-correction edits typically run 10 to 30 per text for human writing and 0 to 10 for LLM output, often near zero. Perfect copy with zero typos, uniform sentence lengths, and high lexical diversity is statistically more likely to be machine-written than a draft with a few rough edges. Bad news for insanely good human copy-editors, to be honest.

Grammar perfection is the most easily defeated signal, since injecting deliberate errors fools error-rate heuristics. And commercial detectors have moved past surface spelling because detectors read deeper stylistic patterns. So treat flawless grammar as one input among several, never as proof on its own.

How AI is homogenizing writing

The tells matter more now because AI is flattening writing across whole platforms, and your brand voice is competing against that flattening.

ChatGPT's release triggered a 189% surge in AI usage in LinkedIn posts. On the broader web, an ACL 2025 study tracked Medium's AI attribution rate rising from 1.77% to 37.03% and Quora's from 2.06% to 38.95% between early 2022 and late 2024, while Reddit grew far slower, from 1.31% to 2.45%.

The stylistic effect is convergence to a mean. A 2025 arXiv preprint concluded that LLMs homogenize writing styles, amplifying dominant characteristics while suppressing individual expression.

And the homogenization is bleeding into humans. A University of Helsinki study comparing essays before and after ChatGPT's release found "delve" and "foster" now used more than ten times as often. A Max Planck analysis of about 280,000 academic YouTube transcripts found AI-associated words like "delve" and "meticulous" rising 40 to 50% in unscripted speech. People are absorbing the model's vocabulary and speaking it out loud.

For a content team, the practical stakes are pretty stark. Early LinkedIn studies suggest a distribution penalty attached to generated content. A study of 3,368 posts found likely AI-generated content received 45% less engagement, and LinkedIn's own detection reportedly cuts reach 20 to 40% on fully AI-written posts. Sounding like the average has become a distribution problem.

How AI writing tells fit into detection

Detection tools formalize these tells into a score, and the scores are less reliable than vendors claim. Turnitin advertises 98% accuracy and under 1% false positives, but only for documents where at least 20% of content is flagged, and independent testing puts false positive rates at 5 to 12% on edge cases like non-native English and heavily edited drafts. A Stanford-affiliated study found seven detectors classified 61.22% of TOEFL essays by non-native speakers as AI-generated. Vanderbilt disabled Turnitin's detector over accuracy and bias concerns.

The gap between vendor claims and independent results is wide across the field. DetectGPT claims 99% but measured 54.63% in a peer-reviewed study, described there as "virtually no better than random guessing." GPTZero performed far better in the same study at 97.22% accuracy with a 0% false positive rate, so the tools are not interchangeable, but no score should carry a decision on its own.

The non-native speaker bias is the reason editorial judgment beats automated scores for anyone managing real writers. The same statistical properties that flag AI, low perplexity and consistent prose, also describe skilled non-native English writing. ESL false positive rates ran 14.4% against 4.15% for published human texts, more than three times higher. Multiple peer-reviewed studies advise against using detection tools as the sole basis for any consequential decision. Use them as a flag that prompts a closer read, not as a verdict.

How to edit out AI tells and restore brand voice

Editing AI copy back into brand voice runs in four passes, in order. It's the sequence we run on our own drafts, and the order matters because the surface work is fast while the voice work decides quality.

Strip the surface tells. Run the vocabulary checklist and cut the marker words. Break the em dash habit, deflate the colon-heavy headers, and flatten the suspiciously symmetrical lists.
Reinject rhythm and voice. Vary sentence length on purpose and let a fragment land a point. Then add the thing the model can't supply: a stated opinion, a specific example only your company would use, a consistent first-person stance that reacts to the material.
Fact-check every claim. Verify statistics and dates against a source. Hallucination rates on recall benchmarks run as high as 48%, so treat any confident, specific factoid as unverified until you've checked it.
Calibrate to the persona. Read the draft against who it's for. Does it sound like your brand talking to this buyer, or like a model talking to everyone? Cut the sycophancy, the forced enthusiasm, the hedges that commit to nothing.

Content teams hit the bottleneck at memory, because most content operations run the strip-and-reinject cycle from scratch on every piece. The AI writing the draft knows nothing about the company, so you re-explain positioning, re-enter competitive context, and re-calibrate voice every session. The output is generic because the input is generic, and no amount of prompt engineering fixes a system with no memory.

GrowthX designed GrowthOS, its Growth Operating System, around that gap. During onboarding, the team builds the Context layer first, mapping competitors and extracting personas from real data, then calibrating voice so every downstream agent reads from it. Voice persists across drafts instead of resetting each session.

Nothing ships without human approval, which keeps the stated-opinion, specific-example work in human hands where it belongs. If you're weighing whether to consolidate a stack of disconnected drafting and editing tools into one operated system with persistent voice, you can book a demo. Engagements start from $6,000/mo.