What is the fastest way to tell if ChatGPT wrote something?

Scan for three things in the first paragraph: em-dash overuse, the word delve, and uniform sentence length. If two of the three are present in 150 words, ChatGPT is the most likely author. Add a high score from a detector and a polite opening like Certainly! and the case is almost decided.

Does ChatGPT really use em-dashes that often?

Yes, more than nearly any other writer. Independent samples of GPT-4 output average two to four em-dash characters per 500 words, compared to fewer than one in typical student writing. The pattern persists across GPT-3.5, GPT-4, and GPT-4o. It is one of the clearest single-character fingerprints in the model.

Why does ChatGPT say delve so much?

Delve is rare in casual English but common in academic and African English usage. RLHF training data and the helpful, formal register that OpenAI optimized for pushed delve into ChatGPT's high-probability vocabulary. By 2024, delve was so overused that researchers used its frequency alone as a weak AI signal.

How is Claude different from ChatGPT in style?

Claude tends to write with more self-correction and fewer fixed cliches. It hedges thoughtfully rather than reflexively, often acknowledging multiple sides without dropping into ChatGPT's symmetric on the other hand pattern. Claude still uses uniform paragraph length and structural lists, which keeps its burstiness low.

How is Gemini different from ChatGPT in style?

Gemini leans into enumerated lists, tables, and bolded headings even in short answers. It often opens with a one-line direct answer, then bullets. Where ChatGPT pads, Gemini scaffolds. Both share low burstiness and repetitive vocabulary, but Gemini's structural fingerprint is more visual.

Can newer GPT models hide these tells?

Partly. GPT-4o is less obvious than GPT-3.5 and avoids some of the worst cliches when prompted carefully. But em-dash overuse, polite openings like Certainly!, and uniform paragraph length persist even in GPT-4o by default. Custom system prompts can suppress some tells but rarely all of them.

Are these tells reliable on academic or technical writing?

Less reliable. Academic abstracts, legal memos, and technical documentation share many features with AI text: formal tone, hedging, careful symmetry, low personal voice. A high score on those genres is more likely a false positive. Always weigh genre before deciding ChatGPT wrote a document.

Do AI detectors check for these specific tells?

Modern detectors combine statistical signals like burstiness and perplexity with vocabulary and structural heuristics that mirror the 12 tells. Our own AI Detector flags the same six clusters teachers use. No detector relies on a single tell. Combining tells is what produces reliable verdicts.

Will rewriting fix the tells?

Yes if you rewrite by hand and add specifics. Replace em-dash characters with periods or commas, swap delve for a verb you actually use, vary sentence length, and drop one personal example per paragraph. Asking ChatGPT to make this sound more human rarely removes the deep statistical fingerprint.

Are these tells biased against non-native English writers?

Some are. Liang and colleagues at Stanford found that GPT detectors flagged 61% of TOEFL essays by non-native English speakers as AI-generated. Formal vocabulary, careful symmetry, and hedging are common in second-language academic English. Use these tells as signals to investigate, never as proof on their own.

ChatGPT Writing Style Tells: How to Identify AI Text in 30 Seconds

Quick Answer

ChatGPT writing carries 12 specific tells: em-dash overuse, polite openings like Certainly!, the words delve and tapestry, navigating the complexities, in today's digital age, hedging phrases, uniform paragraph length, list overuse, symmetric arguments, in conclusion endings, and an overly polite tone. Two or three together in a paragraph is enough to identify GPT output in roughly 30 seconds.

Most readers can learn to recognize ChatGPT in a single paragraph after seeing the pattern named once. This guide lays out the 12 specific signatures, why they appear, how Claude and Gemini compare, and where the tells mislead. To test a paragraph against the same signals teachers use, paste it into our AI Detector tool.

Why ChatGPT Has a Style Signature

ChatGPT writes the way it does for a reason. OpenAI trained the model on a broad slice of internet and book text, then fine-tuned it with Reinforcement Learning from Human Feedback (RLHF). Human raters consistently preferred responses that were polite, balanced, structured, and hedged. The model learned to write that way by default.

The same RLHF preference for helpful and harmless output trained the model to use safe vocabulary, careful symmetry, and a polite register. The result is text that reads competently but predictably. The fingerprint shows up across GPT-3.5, GPT-4, and GPT-4o, with each newer version slightly less obvious but never invisible.

The 12 ChatGPT Tells

1. Em-Dash Addiction

ChatGPT loves the em-dash character. Independent samples of GPT-4 output average two to four em-dashes per 500 words. Typical undergraduate writing has fewer than one. Look for parenthetical phrases set off with em-dashes where a comma or period would do the job. Example: The project was delayed, set off with em-dashes around despite the team's best efforts, until late August.

2. Opening With Certainly! or I'm Here to Help

ChatGPT cannot stop being polite. When given a task, it often opens with Certainly!, Of course!, Absolutely!, or I'm here to help. Even when the prompt asks for a draft and not a chat reply, residue of the assistant register slips in. A formal essay that opens with a one-line affirmation is almost always a paste-from-ChatGPT.

3. Delve Into

The verb delve is rare in casual English. By 2024 it was so over-represented in ChatGPT output that researchers used its frequency alone as a weak AI signal. If a 500-word piece uses delve into once, it could be coincidence. Twice is suspicious. Three times is a fingerprint.

4. Tapestry Of

A close cousin of delve. ChatGPT reaches for tapestry to describe anything complex or interconnected: a tapestry of experiences, a tapestry of cultures, a rich tapestry of voices. The metaphor is functional but the model leans on it far harder than human writers do.

5. Navigating the Complexities

The phrase navigating the complexities (or navigating the challenges, navigating the nuances) appears in roughly one in twenty ChatGPT essays on social topics. It is a hedge that signals depth without committing to a specific claim. Native human writers usually pick a concrete verb instead.

6. In Today's Digital Age

Almost any prompt about technology, society, or culture triggers in today's digital age as an opener. Variants include in our increasingly connected world and in the modern era. The phrase carries no information. It exists to soften the model into its opening paragraph.

7. Hedging Language

ChatGPT hedges constantly. It could be argued that, one might say, some experts believe, this could be seen as. The hedging is RLHF residue: human raters preferred answers that did not commit to strong claims. The result is prose that sounds careful but rarely takes a position.

8. Uniform Paragraph Length

ChatGPT paragraphs cluster around three to five sentences each. Open a five-paragraph response and measure: most paragraphs will land within one sentence of each other. Human writers swing from one-line paragraphs to ten-line ones depending on emphasis. The metronome rhythm of ChatGPT is one of the loudest non-vocabulary tells.

9. List Structures

When in doubt, ChatGPT bullets. Even prose responses are interrupted by numbered or bulleted lists, sometimes for three items that would read better as a single sentence. Gemini is worse on this dimension, but ChatGPT is still far above human baseline.

10. Symmetric Arguments

Every claim gets a counter. On one hand, on the other hand, while X is true, Y must also be considered. The symmetry is so consistent that researchers have used balance-of-argument as a model fingerprint. Human writers more often pick a side.

11. Closing With In Conclusion

A real human essay rarely uses the literal phrase in conclusion. ChatGPT uses it as a default closing transition. Variants include to sum up, in summary, ultimately. The closing paragraph then restates the introduction rather than adding new insight.

12. Overly Polite Tone in Every Register

Whether the prompt asks for a snarky tweet, a formal cover letter, or a casual blog post, ChatGPT defaults to a polite, neutral, professional register. Genuine snark, sharp opinion, and unhedged frustration are rare in default output. The flatness is the giveaway. Real writers have moods. ChatGPT has one mood.

ChatGPT vs Claude vs Gemini: Style Differences

The big three frontier models share a low burstiness baseline and vocabulary repetition, but each has its own fingerprint.

ChatGPT: Verbose, em-dash heavy, hedging-prone, polite openings, delve and tapestry, in conclusion endings.
Claude: Thoughtful pacing, fewer fixed cliches, more self-correction (e.g., actually, on reflection), still uniform paragraph length, prefers commas over em-dashes.
Gemini: Enumerated, table-heavy, list-driven, often opens with a one-line direct answer then bullets, strong structural fingerprint.
All three share: Low burstiness, vocabulary repetition within 500 words, symmetric arguments, generic examples.

Detecting GPT-4 vs GPT-3.5

GPT-3.5 wears its tells on its sleeve. Five-paragraph essay structure, two delves, four em-dashes, and an in conclusion ending appear in a typical 500-word draft. GPT-4 is harder. Em-dash use drops slightly, hedging is more nuanced, and cliches appear less frequently. GPT-4o softens the polite openings further.

But the deep statistical signal stays. Burstiness stays low. Paragraph length stays uniform. Em-dash use stays elevated. Certainly! still slips in. Researchers running benchmark detection on GPT-4o (Mitchell et al. and follow-up work in 2024) found that the strongest detectors still flag GPT-4o text well above human baseline. The tells migrated, but they did not vanish.

Live Demo: Run This Text Through Our Detector

Here is a 110-word sample written by GPT-4 from the prompt write a short paragraph about remote work:

In today's digital age, remote work has fundamentally transformed the way professionals navigate the complexities of their careers. While it offers a tapestry of benefits, including increased flexibility and the ability to delve into a healthier work-life balance, it also presents challenges, such as feelings of isolation and difficulty separating personal and professional life. Companies must adopt robust frameworks to support their distributed teams, leveraging communication tools and fostering a culture of trust. Ultimately, the future of work is ever-evolving, and organizations that embrace this shift while addressing its complexities will thrive in an increasingly interconnected world. In conclusion, remote work is here to stay.

Paste that paragraph into our AI Detectorand the verdict comes back Almost Certainly AI in under a second. The detector flags six cliche phrases (in today's digital age, navigating the complexities, tapestry of, delve into, robust framework, leveraging, ever-evolving, in conclusion), low burstiness (sentences cluster between 18 and 28 words), and structural symmetry (introduce, list benefits, list challenges, conclude). Each is a tell from the list above.

When These Tells Are Misleading

The 12 tells produce false positives in three notable genres.

Academic abstracts. Journal abstracts are formal, hedged, symmetric, and packed with safe vocabulary. They look like ChatGPT because they are written to the same constraints: brevity, balance, and neutrality.
Technical and legal documentation. Specifications, policy briefs, and legal memos prize hedging and symmetry on purpose. A high AI score on a contract is usually a false positive.
Non-native English writers. Liang et al. (Stanford 2023) found that GPT detectors flagged 61% of TOEFL essays by non-native English speakers as AI-generated. Formal vocabulary, careful symmetry, and hedging are characteristic of second-language academic English.

The 12 tells are most reliable on student essays, blog drafts, social posts, marketing copy, and personal correspondence. Use them as signals to investigate, not as proof on their own. If you are a teacher or editor, combine multiple tells before drawing a conclusion. If you are a writer worried about false positives, read our companion guide on how to humanize AI text for the techniques that fix the underlying signal, not just the surface words.

The 30-Second Check

When you suspect ChatGPT wrote something, run this in order.

Count em-dash characters in the first 200 words.
Scan for delve, tapestry, navigating, in today's digital age.
Measure paragraph length variance. Tight cluster around 3 to 5 sentences is a signal.
Look for symmetric arguments and the phrase in conclusion.
Paste the text into our free AI Detector to score the same signals automatically.

Two or three matches in 30 seconds is enough to identify ChatGPT output reliably. Add a high detector score and the case is essentially closed.