14 signals that betray AI-written text
AI text has forensic fingerprints. Em-dashes, tricolons, hedge words, robotic rhythm. Here are the 14 signals that give it away.
AI-generated text looks right at first glance. The grammar is correct, the structure is logical, the vocabulary is appropriate. But read a few paragraphs and something feels off. It is too smooth. Too even. Too polished in a way that real writing never is.
That feeling has a source. AI text carries forensic fingerprints that trained readers and automated detectors can pick up. These are not random quirks. They are systematic patterns baked into how language models generate text.
Here are the 14 signals, with before and after examples for each.
1. Em-dashes as dramatic pauses
AI loves the em-dash. It uses it constantly as a dramatic pause, a parenthetical, or a way to tack on extra information. Human writers use em-dashes very occasionally. AI uses them in nearly every paragraph.
AI version: The results were clear - and somewhat surprising - the model outperformed expectations.
Human version: The results were clear. The model outperformed expectations, which surprised us.
2. Smart quotes and special characters
Language models produce typographic "smart quotes" and special Unicode characters that most humans never type. Your keyboard produces straight quotes. AI produces curly ones, along with special spaces and other invisible characters.
AI version: "It's a 'paradigm shift' in the industry."
Human version: "It's a 'paradigm shift' in the industry."
The difference is subtle visually but obvious in the raw character codes.
3. Uncontracted forms
AI defaults to formal, uncontracted language. "It is" instead of "it's." "They are" instead of "they're." "Do not" instead of "don't." Real writing, especially anything informal, uses contractions constantly.
AI version: It is important to note that they do not have access to the full dataset. We should not assume otherwise.
Human version: It's worth noting they don't have access to the full dataset. We shouldn't assume otherwise.
4. Formulaic transitions
AI reaches for the same transition words every time. "Furthermore." "Moreover." "Additionally." "In conclusion." These words exist in English, but human writers use them sparingly. AI uses them as crutches between every paragraph.
AI version: Furthermore, the data suggests a correlation. Moreover, this aligns with previous research. Additionally, we observed similar patterns in the control group.
Human version: The data suggests a correlation, and it lines up with previous research. We saw similar patterns in the control group too.
5. Tricolons (groups of three)
AI structures lists and examples in groups of three with suspicious consistency. Three adjectives, three examples, three benefits. Real writing varies. Sometimes you list two things. Sometimes five. AI defaults to three.
AI version: The tool is fast, reliable, and intuitive. It handles documents, images, and spreadsheets. Users praise its speed, accuracy, and simplicity.
Human version: The tool is fast and reliable. It handles most document types and a few image formats. Users mostly praise the speed.
6. Passive voice overuse
AI leans on passive constructions. "The document was processed" instead of "we processed the document." Passive voice has its place, but AI uses it to avoid taking a position or assigning agency.
AI version: The file was uploaded and processed. Errors were identified and corrections were applied automatically.
Human version: You upload the file, the tool processes it, and it flags errors for you to fix.
7. Hedge words
AI hedges constantly. "It is worth noting," "it should be mentioned," "one might argue," "it is important to consider." These phrases add nothing. They are filler that makes the text feel cautious and non-committal.
AI version: It is worth noting that this approach might potentially offer some advantages. One could argue that the benefits are significant.
Human version: This approach has clear advantages.
8. Uniform sentence length
Read any AI paragraph and count the words per sentence. They cluster around the same length. Real writing has short punchy sentences mixed with longer ones. AI produces a steady, metronomic rhythm.
AI version: The system processes documents in real time. It analyses the content using advanced algorithms. Users can configure the settings to match their needs. The output is generated within seconds of submission.
Human version: The system processes documents in real time. Fast. You configure what you need, and it spits out results in seconds.
9. Topic-sentence-first paragraphs
AI almost always opens a paragraph with a topic sentence that summarises what follows. Every paragraph reads like a textbook. Real writing sometimes leads with a detail, an anecdote, or a question before getting to the point.
AI version: Privacy is a fundamental concern in modern web development. Users expect their data to be protected. Developers must implement robust security measures.
Human version: Your users do not want their files uploaded to a server they have never heard of. Start there. Build the privacy model around that constraint.
10. Overused vocabulary
AI has favourite words. "Delve," "tapestry," "landscape," "nuanced," "multifaceted," "comprehensive," "robust." These words appear in AI text at rates far above normal human usage. If you see "delve" in a blog post, it was probably generated.
AI version: Let us delve into the nuanced landscape of modern privacy. This multifaceted tapestry of regulations presents a comprehensive challenge.
Human version: Privacy rules are complicated. Here is what actually matters.
11. Colon-heavy structure
AI uses colons to introduce everything. Lists, explanations, examples, definitions. Real writing uses colons occasionally. AI treats them as a universal connector.
AI version: The benefits are clear: speed, privacy, and reliability. The approach is simple: process everything locally. The result: no data ever leaves your device.
Human version: The benefits are obvious. It is fast, private, and reliable. Process everything locally and your data stays on your device.
12. Robotic rhythm
AI text has a consistent, almost musical cadence. Read it aloud and you will hear it. Every sentence follows a similar stress pattern. Real writing stumbles, varies, speeds up and slows down. AI maintains a steady beat.
AI version: The platform offers powerful features. The interface provides intuitive navigation. The system delivers reliable performance. The tool ensures complete privacy.
Human version: The features are solid. Navigation is fine, nothing fancy. Performance has been reliable so far. And your data goes nowhere.
13. Missing contractions in casual context
This is related to signal 3 but specifically about mismatched register. AI writes "cannot" and "will not" even in blog posts, social media, and casual emails where a human would always contract.
AI version: You will not need to worry about data collection. We do not store anything on our servers.
Human version: You won't need to worry about data collection. We don't store anything.
14. Perfect parallel structure
AI constructs parallel lists with mechanical precision. Every item follows the exact same grammatical pattern. Real writers vary their list structures, sometimes intentionally, sometimes just because they are human.
AI version: Our tool processes PDFs quickly, compresses images efficiently, compares text accurately, and cleans HTML thoroughly.
Human version: Our tool handles PDFs, compresses images, and does text comparison. The HTML cleaner is newer but works well.
How we score these signals
Unwrite GPT uses a heuristic scorer that checks for all 14 of these signals. Each signal contributes to a probability score between 0 and 1. A score near 0 means the text reads as human. A score near 1 means the text is saturated with AI fingerprints.
The scorer is not a binary detector. It does not say "this is AI" or "this is human." It measures how many of these patterns are present and how strongly. A human writer who happens to use a lot of em-dashes will score slightly higher on that signal but low on the others. AI text lights up across most of them.
The signals are weighted by how reliably they distinguish AI from human writing. Em-dash overuse and hedge words are strong signals. Smart quote usage alone is weak. The combination is what matters.
What to do about it
If you use AI to draft content, run it through Unwrite GPT. It addresses all 14 signals: normalising punctuation, varying sentence length, replacing formulaic transitions, adding contractions, and breaking up the mechanical rhythm.
You can also use our text comparison tool to see exactly what changed between your AI draft and the humanised version. Side by side, the differences are obvious.
The goal is not to hide AI usage. It is to produce text that reads naturally, engages real readers, and does not trigger every detection heuristic on the internet.