9 signs of AI writing: how to spot AI-generated content in 2026

Updated 21 October 2026. The nine signs now open the page as a quick list, with a rule for when they add up. New: a third false positive (non-native English writers), a section on checking a draft with free tools, and a corrected example of a famous dash-heavy writer.

The signs of AI writing are habits language models fall into by default: too many em dashes, a small set of stock words, the same two-part hooks, evenly weighted bullets, a generic closing question, and prose that is fluent but says nothing you would remember. One sign on its own proves nothing. Three or more in a short piece is a strong signal that a model drafted it or heavily edited it.

The 9 signs a piece of writing was written by AI:

• 1. Em dash density. Two or more em dashes in a paragraph under 100 words.

• 2. The AI vocabulary cluster. Delve, leverage, unlock, navigate, elevate, foster, harness, robust, seamless.

• 3. The symmetric hook. “Most people think one thing. The reality is another.”

• 4. The “not just this, but that” frame, used again and again.

• 5. The beige bullet middle. Four or five evenly weighted bullets that could sit in any post.

• 6. The generic closing question. “What’s your take?” instead of a specific one.

• 7. Even paragraph rhythm. Same-length paragraphs, no fragments, no long tangents.

• 8. Voice-flat coherence. Fluent, correct and forgettable.

• 9. Missing taboos. Every sharp edge softened into a balanced both-sides hedge.

How to use the list: read for the cluster, not a single mark. If you want a second pair of eyes, VoiceMoat’s free AI detector highlights the exact words and patterns in your text that read as AI, each with a human fix, and runs in your browser with no sign-up. The rest of this page explains each sign, the three false positives that catch out lazy readers, and the byline test that catches what the list misses.

Why this matters now: AI-assisted writing is common on every major platform, so the question is no longer whether a given post might be AI. It is whether you can tell, and whether your own writing has drifted into the same patterns without you noticing. The signs work both ways: as a reader’s filter and as a writer’s audit.

Sign 1: em dash density

The em dash is the best-known sign of AI writing because the major models use it as a default rhythm tool, at a rate few writers match outside long-form essays. Two or three in a single short post. Em dashes mid-clause, mid-sentence, mid-list, where a comma, colon or full stop would do the same work with less drama. The habit shows up on Twitter, LinkedIn, blogs and email regardless of topic, because it belongs to the model rather than the writer.

A rule of thumb, not a law: count the em dashes in any one paragraph. Two or more in under 100 words raises the odds of AI help sharply. Three or more makes it very likely, unless you know the writer is an essayist whose voice naturally runs dash-heavy.

Sign 2: the AI vocabulary cluster

Certain words turn up in AI-drafted writing far more often than in human writing. The list shifts from model to model, but a stable core has held through 2026: leverage (as a verb), delve, unlock, navigate (in the metaphorical sense), elevate, foster, harness, robust, seamless, comprehensive, holistic, “in the realm of” and “in the world of”. None of these words is wrong on its own. The sign is the cluster: three or four in one short post. The full list, with plain-English swaps, is in the words AI overuses (and how to ban them).

Sign 3: the symmetric hook

AI-drafted hooks lean on symmetric two-part openings. “Most people think one thing. The reality is another.” “It is not about the tactic. It is about the system.” “Forget the hacks. Focus on the basics.” The structure is fine in moderation. The sign is frequency: when a creator opens four posts out of five with the same two-part pattern, the drafting tool is shaping the feed. Why models do this is explained in why AI drafts read the same, and the practical fix is in how to make AI writing sound human.

Sign 4: the “not just this, but that” frame

One construction deserves its own line: “It is not just a tool. It is a mindset.” Or its sibling: “This is not about followers. It is about trust.” Models reach for it to add apparent depth to a flat statement. Used once, it is a fine rhetorical move. Used three times in one post, it is a fingerprint.

Sign 5: the beige bullet middle

Long AI-drafted posts tend to collapse into a bulleted middle where every bullet is the same length, starts the same way, and says something true but unspecific. “Build trust with your audience. Show up consistently. Provide value with every post. Engage authentically.” Each line passes a quick read. Together they are the sign. Human bullet lists tend to be uneven (one long, one short, one a fragment) and usually include at least one bullet only that writer would write.

Sign 6: the generic closing question

AI-drafted posts default to the same kind of ending: “What’s your take?”, “Drop your thoughts below.”, “Curious to hear your experience.” The sign is the genericness, not the call to action itself. A specific question (“who else has shipped on a Friday and regretted it?”) reads as a person; the generic one reads as a model default.

Sign 7: even paragraph rhythm

Human writers have uneven rhythm. Three short sentences. Then one long, meandering one with an aside. Then a fragment. AI drafts tend toward paragraphs of similar length with similar sentence counts and few fragments, so the shape reads as suspiciously even. It is hard to spot in one post and obvious across a feed: when every post has the same shape, something is normalising it.

Sign 8: voice-flat coherence

The hardest sign to describe and the most reliable once you are tuned to it. The grammar is right, the structure is right, the cadence is right, and there is no detail you would remember tomorrow. Try the substitution test: swap the author’s name for any other creator in the same niche. If the post still reads the same, it is voice-flat. The AI slop essay names the pattern, and the 10 signals of voice explains what is missing.

Sign 9: missing taboos

Real writers have taboos: words they refuse, framings they avoid, opinions they will not soften. Those refusals are part of what makes writing recognisable. AI-drafted writing has none by default. It uses every safe word, includes every conventional framing, and turns every sharp edge into a balanced hedge. If a post is carefully inoffensive on every dimension, the chance a model drafted it is high.

Three false positives worth knowing

Calling human writing “AI” is its own credibility problem. Three groups get misread most often.

• Dash-heavy essayists and poets. Some writers use dashes at rates that look AI-shaped if you only count punctuation. Emily Dickinson built whole poems on the dash, more than a century before any language model. The cluster test clears them: if the vocabulary, hooks and opinions are unmistakably their own, the dashes are voice, not a sign.

• Writers who use AI only to edit. A creator who drafts in their own voice and runs the draft through a model for grammar can end up with dashes and a few stock words added by the edit. Some signs fire (punctuation, vocabulary) while others do not (taboos present, specific details kept, hooks idiosyncratic). That is mixed authorship, not an AI post.

• Non-native English writers. Plainer vocabulary and more regular sentence patterns can look “AI” to readers and to software. Researchers found that popular AI detectors flagged essays by non-native English speakers as AI far more often than essays by native speakers (Liang et al., 2023). Read for specifics and opinions before you judge.

Should you stop using em dashes to avoid looking like AI?

No, and this is the most common overcorrection of 2026. The dash is a centuries-old mark that good writers used for emphasis, interruption and asides long before any model existed. The sign is never the single mark; it is density layered with the rest of the cluster. Cutting a mark you genuinely use, just to dodge a detector, is the same mistake as flattening your vocabulary to dodge one: you remove a real voice signal and end up reading more generic, not less artificial.

The honest rule is proportion, not prohibition. If dashes are part of how you write, keep the ones that do work and cut the ones a model would have inserted for rhythm. If they are not part of your natural register (and for most short-form Twitter writers they are not), the cleaner default is zero, because every em dash in a short post is now read against the AI base rate. The deeper case for protecting voice rather than sanding it down is in authenticity as a moat.

The byline-removal test

The single most reliable check: remove the byline and ask whether you would still know who wrote it. If yes, the writing carries voice signals strong enough that a general model is unlikely to have produced it unaided. If no, the post is either AI-drafted or written in a register so generic that the AI-versus-human question barely matters. The test captures the real question, which is not “was this AI?” but “does this writing belong to a specific person?”

How to check a draft for the signs of AI writing

Reading for the nine signs takes about 30 seconds per post once you are practised. Three ways to speed it up or double-check it:

• Highlight the patterns. The free VoiceMoat AI detector marks the exact words and patterns that read as AI, groups and counts them, and gives a human fix for each. It does not give a “percent AI” verdict, and nothing you paste is uploaded.

• Rewrite the flagged parts. The free Humanize AI Text tool rewrites AI-sounding text toward a natural voice (yours, if you add a sample of your posts) and shows a count of the AI tells it removed. Treat the output as a first edit, then add the specifics only you know.

• Know what scoring detectors can and cannot tell you. Tools like GPTZero and Originality.ai return a probability, and every one of them misfires on some human writing. Our comparison of the best-known AI detectors covers what each catches, misses and costs.

What to do about it as a writer

If you use AI anywhere in your workflow, the nine signs are your audit checklist. Run them on your last 10 posts: count the em dashes, scan for the vocabulary cluster, read your hooks for the two-part pattern, check your bullets for evenness, and look at whether your closing lines are all the same generic question. It takes about 10 minutes and shows where AI polish has flattened your voice.

The fix is not to stop using AI. It is to use AI in a workflow where your voice is a constraint, not an afterthought.

• While drafting: how to avoid the AI tells turns each sign into a practice, with before-and-after examples.

• Over time: voice drift covers the slow slide that creeps in even when you are careful.

• Your audience’s view: can your audience tell you’re using AI? covers who notices and whether they care.

• Choosing an assistant: how to tell if AI really writes like you is a 7-test check for any tool, including ChatGPT and Claude.

Where audenAI fits

audenAI, the brain inside VoiceMoat, starts from the opposite end of this list. VoiceMoat is a Personal Brand OS for LinkedIn and Twitter: audenAI learns from your own posts, so drafts in Studio begin from your patterns rather than the model’s defaults. Words like “leverage” and “delve” can go in Your rules, and it will not use them. Drafts are checked for Style and Stance: instead of asking “does this read as AI?”, the check asks “does this read like me?”, which is the byline test in product form. More on that in why VoiceMoat checks Style and Stance.

Quick checklist

• Em dash density: two or more in a paragraph under 100 words.

• Vocabulary cluster: three or more of leverage, delve, unlock, navigate, elevate, foster, harness, robust, seamless, comprehensive, holistic.

• Symmetric hook: “Most people think one thing. The reality is another,” repeated across the feed.

• The “not just this, but that” frame: fine once, a fingerprint three times in one post.

• Beige bullet middle: four or five evenly weighted bullets that could appear anywhere.

• Generic closing question: “What’s your take?” rather than a specific question.

• Even rhythm: no fragments, no uneven sentence lengths.

• Voice-flat coherence: fluent, forgettable, swappable byline.

• Missing taboos: no refusals, no sharp edges, a careful hedge on everything.

• Finally, run the byline-removal test.