1. Heavy em-dash use
Some models use the long dash (—) noticeably often. Claude, Gemini, and ChatGPT can all be instructed to drop the habit by now. Microsoft Word sometimes auto-replaces "-" with "—" on its own. I.e. never judge it alone.
!
In short
Certain language patterns show up disproportionately often in AI-generated text: heavy em-dash use, template phrases like "this isn't just X, it's Y", uniform sentence length, a noticeably cautious and neutral tone, unusually predictable vocabulary, and the absence of small personal imprecisions. None of these is proof on its own - all of them can be deliberately avoided. They're a first signal that justifies a second check, not a replacement for one.
Why look at language patterns at all?
A detector isn't always within reach; a trained eye is. Language models like ChatGPT write more statistically predictably than most people do - that produces recurring, nameable patterns. The key difference from detector software: these patterns are consciously avoidable once someone knows what you're looking for. That's why this is the first step, never the last.
At PlagiatGPT we've analyzed millions of documents to train our own detector - and we can say with confidence: identifying sentences and tells like these alone can never build a reliable AI detector. It's actually worse than that: our own models measurably got worse whenever they leaned too heavily on exactly these surface-level tells.
In practice, that means
The style of a text is definitely a hint - but not more than that. Newer language models increasingly try not to sound like AI, and students frequently rephrase their own text before handing it in, too. The six tells below stay useful anyway - as a first, quick look, not the final word.
SIX TELLS FROM THE FIELD
The 6 tells in detail
2. The "it's not just X, it's Y" sentence
This template shows up markedly more often in AI text than in comparable human writing. One hit is coincidence; several in the same text isn't.
3. Even sentence length
This is "burstiness" in text form: human writing swings between short and long sentences, AI text tends toward a more even rhythm. Read it aloud - an unusually flat pace stands out faster than you'd expect.
4. A noticeably cautious, neutral tone
No clear stance, no sharp claims, every statement immediately hedged. A warning sign for a paper that's supposed to argue a position - less telling for a plain summary.
5. Unusually predictable vocabulary
This is "perplexity": AI text often picks the most obvious word, rarely an unexpected one. Reads smoothly, but rarely like a personal voice.
6. No small personal imprecisions
Real writing by students often has small, specific details, detours, or even small mistakes. A flawless but impersonal text with no personal anecdote or "color" at all is more conspicuous than it looks.
🛈
The bottom line: every one of these six tells can disappear with a different prompt, a different model, or a short revision pass. They're early hints of a possible "one-shot" (copying an AI's output for a prompt 1:1) - not proof. The next step is always a second, documented check. Never make a decision based on this list alone.
The most common trap: just asking a chatbot
A text looks suspicious, so plenty of lecturers and graders paste it straight into ChatGPT, Claude, or Gemini and ask: "Is this AI-generated?" The answer sounds convincing, but in practice it's worthless. General-purpose AI chatbots were never trained to tell AI text apart from human text - that's a different task from generating text.
Our tip: run the first suspicion past an actual (free) AI detector
A purpose-built AI detector is fundamentally different from a chatbot you ask for its opinion: it was trained specifically to tell AI-generated text apart from human writing - and it's far more accurate than any human could ever be by eye, no matter how practiced. So don't let a chatbot "make up" your first suspicion for you - use a tool built specifically for this instead. That doesn't replace a second, documented check for a serious case - but it's the right first step.
Frequently asked questions
No. The em dash is a hint - plenty of people use it naturally, and newer ChatGPT versions can be instructed to avoid it. Never judge a single tell in isolation.
Yes, fairly easily - with the right prompt or a short revision pass. That's exactly why these six are a first signal, never the last step of a review.
Use an AI detector to confirm the suspicion.
By Simon Schmidlin
Keywordsdetect chatgpt text, ai text detection, signs of ai-generated text, chatgpt text in academic work, ai writing red flags