Want to know how to tell if something is written by AI? Look for flat rhythm, safe buzzwords, no personal stories, and facts that fall apart under a nudge. Then run the text through an AI checker and read the score as a signal, not a verdict. No single clue is proof on its own.
That last part matters. One tell means nothing. A stack of tells, plus a checker score, plus context about who wrote the piece, adds up. Here is how to make sense of the AI writing itself, and of the AI detectors, before you trust either.
The 60-second gut check
Skim the piece and ask a few questions. You’re not scoring yet, notice what the writing does.
- Do the sentences all feel the same length, with a smooth writing flow?
- Are there safe phrases and buzzwords where a real opinion should be?
- Are there bullet points and tidy lists everywhere, even where they don’t fit?
- Is there a single specific detail? A name, a date, a scar?
- Does the tone stay perfectly level, start to finish?
If you nodded four or five times, you are probably staring at AI writing. Now, the eight tells.
8 signs something was written by AI
1. It leans on safe buzzwords and empty phrases
Generative AI plays it safe. The AI reaches for the phrases that show up most in its training data, so you get “in today’s landscape” and “plays a crucial role” where a person would be blunt. The writing reads fine, it never commits. That hollow, agreeable style is the clearest tell in AI content.
2. Every sentence is the same length
Read a few lines aloud. Human writing has a pulse. Short punch. Then a longer sentence that doubles back on itself. AI generated writing flattens that into one rhythm, similar sentence length line after line, a metronome, not a heartbeat. When the sentence structure never varies and the writing flow feels too smooth, that is a tell.
3. There’s depth on the surface but no real ideas
The AI can create a page of text and say nothing. What AI content rarely does is pick a side, chase a tangent, or land on ideas you haven’t read before. Plenty of words, thin center. If you can’t say what the writer thinks, that’s your answer.
4. No personal stories, no specifics, no mess
Real humans leave fingerprints, and so does real human writing. The night your code broke at 2 a.m. The teacher who ruined poetry for you. AI-generated text skips all of that. The AI generalizes. Humans write “my landlord kept the deposit, so I photograph every wall now.” AI writes “documentation matters when renting.” One of those has a heartbeat.
5. The facts and citations don’t hold up
A large language model predicts likely-sounding text, so it pulls a citation out of thin air, confident and wrong. Check a couple of claims, click the links. If the sources are vague, misattributed, or point to studies that don’t exist, that is a strong tell the text is AI-generated. Human writers get things wrong, but they leave a checkable trail.
6. The tone stays oddly neutral the whole way through
People get annoyed. They overshare and contradict themselves halfway down the page. AI-generated prose keeps an even, agreeable voice start to finish, like support chat that never breaks character. A little friction, a shift in mood, a joke that misses, those read human. Relentless calm does not.
7. The punctuation and formatting have tics
Watch the tidy quirks. Perfectly balanced lists. Bullet points for everything, even where a plain sentence would do. Headings on headings. And yes, the em dash, which generative AI scatters far more than most writers do. None of these prove anything alone. A human can love lists and dashes too. But when the same tics repeat across AI-generated text, add them to the pile.
8. Run it through an AI checker and read the score carefully
Now get a second opinion. Paste the passage into an AI detector, which scans for AI-generated patterns and returns a likelihood score with the flagged sentences highlighted. Treat that number as one more signal. Many tools want a free account first. Students, editors, and teachers all lean on these AI tools now. High score plus five manual tells is a real case. High score alone is not.
How AI detectors actually work
No magic wand. Math and patterns.
What the detector is really measuring
Most AI detectors run your text through a machine learning model trained on piles of human written text and AI-generated text, then guess which one your sample resembles.
A big part of that guess is “perplexity,” how predictable each word is. If the next word is always the obvious word, the AI detector reads the output as machine-like. If your phrasing surprises it, the score leans human. That is why AI detectors work at all, and why AI-generated text fools them easily.
Why solid human writing gets flagged
Here is the uncomfortable part. Clean, predictable human writing looks a lot like AI writing to these tools, even when a person wrote it. A Stanford study ran 91 TOEFL essays by non-native English speakers through seven detectors and found more than half wrongly flagged as AI-generated, one detector flagging nearly all of them, while native-speaking students came back correctly as human. Turnitin’s own guidance is cautious enough that scores from 1 to 19 percent now show only an asterisk, because false positives cluster there. Even honest human written documents get flagged. A flag is not a conviction.
No AI checker is 100 percent accurate
Even the builders admit it. When OpenAI released its own detector, it caught just 26 percent of AI-generated text and mislabeled human writing as AI 9 percent of the time, then shut the tool down for low accuracy.
In the largest independent benchmark to date, RAID, detectors that looked strong collapsed once text was paraphrased or lightly edited. The 2024 NIST GenAI pilot found the same: detection accuracy swings wildly with the tools, and slips as AI models improve. No detector is fully accurate. Every AI checker gives you a probability, not a fact.
| Signal | Reads AI | Reads human |
|---|---|---|
| Rhythm | Even sentence length, glassy flow | Bursts, fragments, uneven pace |
| Detail | General, safe, tidy | Specific names, dates, small messes |
| Stance | Neutral, no real point | An actual opinion, a stake |
| Sources | Confident, sometimes invented | Checkable, sometimes wrong and owned |
Reading an AI checker score without panicking
Picking an accurate AI detection tool helps, but the number still needs a human to read it. Many promise “unmatched accuracy.” Grammarly, for example, claims 99 percent accuracy on the RAID benchmark. That figure is real in the lab, and close to meaningless for your essay, because the same AI detection tools crater on edited AI-generated text. So read any percentage as an accurate signal only when the writing backs it up.
Turn the percentage into a risk level
Think in bands, not absolutes. A low score means little to worry about. A middle score means read the document again for real human writing, not AI-generated filler. A high score is worth a closer look and an honest conversation, not a penalty. The score points you toward a question, never the answer.
Cross-check before you accuse anyone
Run the same document through two or three AI detection tools, since other tools may disagree and none reads accurately. Run plagiarism checks alongside, for example, because a checker can flag potential plagiarism that an AI detector misses. Then weigh the score against the writer’s usual style, drafts, and edit history. For more reliable results, confirm you’re reading human written content, not AI-generated text, before you judge.
Using AI responsibly
If you use AI yourself, a few habits keep you out of trouble and build trust with readers, editors, and instructors.
Disclose it when the rules say so
Check the policy before you submit. Many classes, journals, and clients now ask you to say how you used generative AI tools. A one-line note on what the AI did and what you did is cheap insurance. If you’re a student, it protects you when an assignment gets flagged, essay or blog post.
Keep your drafts and edit in your own voice
Draft somewhere that keeps a record. Google Docs saves your version history, and that trail of edits, rewrites, and comments is the best proof of human authorship you can create. If a detector flags your work later, your Google Docs history and your writing process beat any argument. Then rewrite the AI parts in your own words and voice, so the ideas sound like you, not the model that created them.
Where students and teachers land
In school this gets sharp fast: academic integrity rules can turn one careless assignment into a real problem, and instructors lean on AI detectors to screen student work. If you teach, treat a flag as the start of a conversation about the writing process, not the verdict, and ask the student for their drafts. If you’re one of the students, keep your work in Google Docs, ready to show. Talking it through beats a silent accusation, and keeps academic integrity intact in the AI era.
The bottom line
You can’t catch every AI-generated passage, and shouldn’t pretend to. Read the writing first, let a detector back you up, keep the receipts you created. Then trust the pile of evidence, not the single number.
FAQ
The usual tells cluster together: safe buzzwords instead of opinions, sentences that run the same length, bullet points everywhere, no personal stories, an even tone with no real voice, and facts that don’t check out. One tell proves nothing. Several at once, backed by a checker score, is when to pay attention to AI content.
An AI-generated document usually reads smooth, balanced, and a little hollow. The grammar is clean, the structure neat, and the writing covers a topic without a real point or a concrete example. It rarely surprises you. That polished, generic style is the fingerprint most AI systems leave on their output, no matter which AI models produced it.
The “30% rule” is an informal habit, not an official standard. Some instructors treat a detector score above 30 percent as a reason to look closer at an assignment. No school, and no detector maker, endorses it as a hard line. Even Turnitin says its score should never be the sole basis for a decision, so treat 30 percent as a prompt to review, not proof.
A 40 percent reading means the AI detector thinks a meaningful chunk of the document matches AI-generated patterns, and many schools trigger a closer review. Bad? Not necessarily. False positives are common, especially for clear or non-native human writing. For students, a 40 percent score is a reason to show drafts and talk, not a verdict that anyone cheated.
Sources
- OpenAI, “New AI classifier for indicating AI-written text” (26% true-positive rate; tool withdrawn July 2023 for low accuracy): https://openai.com/index/new-ai-classifier-for-indicating-ai-written-text/
- Liang, Yuksekgonul, Mao, Wu, Zou, “GPT detectors are biased against non-native English writers,” Patterns (Cell Press), 2023: https://www.sciencedirect.com/science/article/pii/S2666389923001307
- Dugan et al., “RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors,” ACL 2024: https://aclanthology.org/2024.acl-long.674/
- Grammarly, AI Detector product page (99% RAID accuracy claim): https://www.grammarly.com/ai-detector
- University of Georgia, Center for Teaching and Learning, “Academic Honesty and Generative AI” (Turnitin false-positive band and ESL caution): https://ctl.uga.edu/academic-honesty-and-generative-ai/
- NIST, “2024 NIST GenAI (Pilot Study): Text-to-Text Evaluation Overview and Results,” 2025: https://www.nist.gov/publications/2024-nist-genai-pilot-study-text-text-evaluation-overview-and-results