How to detect AI generated text

Short answer: you can't spot AI text by hunting single words. Read for three things instead. Sentence rhythm, how neatly each paragraph resolves, and whether any detail could only have come from a person. Flat on all three means read it again slowly.

AI generated text is everywhere now. In your inbox. In blog posts. In student essays. In marketing copy. And unless someone told you upfront that AI wrote it, you probably would not know.

The tools meant to catch AI writing aren't great at this either. Most detectors hover around 70 to 80 percent accuracy on a good day. They flag clean human writing as AI all the time, especially if the writer is a non-native English speaker. And they can be fooled by basic rewrites and paraphrasing.

So if you're an editor, a teacher, a content lead, or just someone who reads a lot of text and wants to know what is real, you need something better than a probability score. You need to know what AI writing actually looks like.

This guide covers 8 practical, manual signs of AI generated text. No paid tools. No browser extensions. Just observation, pattern recognition, and a bit of scepticism.

1. Sentences that are all the same length

Human writing is messy. We write short sentences. Then longer ones. Then a fragment. Then something that runs on because we got carried away. The rhythm shifts constantly because our thinking shifts constantly.

AI writing does not do this naturally. Most language models produce sentences that are roughly the same length, one after another. The pacing is flat. Read a paragraph aloud and if every sentence feels like the same number of beats, that is a signal.

Detectors call this low burstiness. Burstiness is just variation in sentence structure and length. Humans burst. Machines do not, unless someone explicitly prompts them to.

2. Transitions that feel like a template

AI loves certain transition words. Academic-sounding connectors that feel like they came from a textbook index. These words are not wrong. But when they appear in every paragraph, one after another, the text starts to feel like it came off an assembly line.

Real writers use transitions inconsistently. Sometimes we say "also." Sometimes we just start the next sentence with no connector at all. Sometimes we use contrast: "But here is the thing." AI rarely does that unless someone tells it to.

Scan for paragraphs that all start the same way. If every section opens with a predictable connector, the text might have been generated.

3. Everything is too balanced

AI writing tends to present both sides of every argument with equal weight. This sounds fair. In practice, it reads like someone who is scared to take a position.

Human writers pick sides. We exaggerate. We get annoyed. We use words like "honestly" and "the real problem is" and "I have seen this fail too many times." AI writing rarely shows frustration or conviction. It stays level. Always level.

This is one of the strongest tells in longer pieces. If you finish reading 2000 words and have no idea what the author actually thinks, check whether a machine wrote it.

4. The vocabulary is too clean

Human writers repeat words. We use the same adjective three times in a paragraph without noticing. We say "thing" when a more precise word exists. We write "really good" instead of "exceptional." It's not elegant, but it's real.

AI text rarely does this. It picks the most fitting word every time. It avoids repetition aggressively. Every paragraph reads like it was copyedited by someone who was paid by the hour.

If the vocabulary feels too precise, too varied, and too clean, that itself is a signal. Human writing has friction. AI writing is lubricated.

5. No personal stories or specifics

AI can tell you that "many writers struggle with finding their voice." It can't tell you about the time it stayed up until 2 AM rewriting a landing page and realized the problem was not the copy but the offer.

Personal anecdotes, specific situations, real examples with names and dates: these are hard for AI to fabricate convincingly. When a piece of writing lacks any concrete detail, any story, any "here is what happened when I tried this," be suspicious.

Generic writing is the easiest kind of text for AI to produce. If an article about productivity never mentions a specific tool, a specific failure, or a specific person, that absence is telling.

6. Repetitive sentence starters

Open a piece of text and look at the first word of every sentence. If you see the same pattern repeating ("The... The... The..." or "This... This... This..."), that's a red flag.

AI models often get stuck in syntactic loops. They produce sentences that all follow the same grammatical template, even when the content changes. Humans don't do this naturally. We start sentences with "And" and "But" and "So" and "Honestly" without thinking.

This is especially visible in longer paragraphs. Try reading just the first three words of each sentence. If they all sound like chapter titles from a textbook, you are probably looking at AI output.

7. No real-world context or edge cases

AI writing tends to stay at the surface level. It explains concepts clearly. It structures information well. But it rarely goes deeper into how things actually work in messy, real-world situations.

A human writer might say: "This method works unless you are in a regulated industry, in which case legal will flag it." AI rarely adds constraints, exceptions, or conditional reasoning. It presents information as if the world is simpler than it actually is.

Look for paragraphs that anticipate objections. Look for sentences that start with "The catch is" or "One thing people get wrong" or "This breaks down when." These are human moves. AI can fake them, but not consistently across a full piece.

8. The ending is a summary, not a conclusion

One of the easiest places to spot AI writing is the last paragraph. AI loves to end by restating exactly what it already said. It is a recap, not a takeaway.

Humans end differently. We leave the reader with something to think about. We make a call to action. We share a final thought that was not in the introduction. We pick one idea from the piece and push it further.

If the last paragraph could have been written by someone who only read the subheadings, the whole piece was probably generated. A real conclusion adds something. A fake one just reminds you what you already read.

Manual detection isn't foolproof. None of these signs work in isolation. A single short paragraph with clean vocabulary means nothing. But when five or six of these patterns show up together, you are almost certainly reading AI output, even if it has been lightly edited.

The tools will keep getting better. So will the models. But the gap between how humans write and how machines write is still visible if you know where to look. Train your eye on these signals. Read more carefully. And if you want to understand what the detection tools are actually measuring under the hood, check out our breakdown of how AI detectors work, or see how the major tools stack up in our comparison of the best AI detection tools.

Want hard numbers? See how the major detectors scored in our accuracy comparison.

The August 2026 update: what the new detector scores changed

The detector market moved again in August 2026. Most tools now combine the classic probability score with a style score that checks rhythm, connectors, and sentence variety. That means some of the tells in this guide now show up in the output before you read a word.

The good news: the manual signs still work, and in some ways they matter more. A detector gives you a number. These eight signs tell you why the number is what it is, and whether the text was edited after generation. No tool can tell you that.

The 2026 wrinkle is short text. Under about 200 words, style scores stay unreliable, which is exactly where the manual checklist shines. For an email, a comment, or a product blurb, your eyes beat any score.

One more thing the updates did not fix: false positives. Detectors still flag clean human writing, especially formal academic prose and non-native English. Use the signs here first, and treat any tool output as a hint, not a verdict.

Want the numbers on which tools improved and which stayed flat? Our AI detection accuracy comparison retests the majors after the August updates.

For the full method behind the scores, how AI detectors work explains perplexity, burstiness, and the new style signals in plain English.

The five-question test for suspicious text

Before you reach for a detector, run this five-question test. It takes two minutes and works on any sample, including text under 200 words where detectors fail.

Question one: does every sentence run a similar length? Count five sentences. If all five land between 15 and 25 words, rhythm is the tell.

Question two: does every paragraph open with a connector or a setup? Scan the first three words of each paragraph.

Question three: is the ending a summary? If the last paragraph restates the subheadings, that is the too-clean wrap-up.

Question four: are there any specifics only a person would know? A name, a date, a tool, a place. The absence is not proof, but the presence is strong evidence of a human.

Question five: would you say this out loud? If the register is more formal than your own speech, the sample leans machine.

Two yes answers warrant a second look. The detector accuracy comparison and the how detectors work guide explain what the tools add on top.

The five-question test is not a detector. It is a habit, and it is the reliable skill in 2026.

One correction to the common advice. If you are reading for words like “furthermore” and “moreover,” you are reading the wrong layer. By September 2026 the models stopped leaning on those as hard. The reliable tells live one level up, in rhythm and in how the argument moves.

The rhythmic tell is a steady pulse. Human writing has beats of different lengths, short punchy sentences next to long winding ones. Machine text tends to arrive in even, numbered units that all breathe the same way.

The argument tell is that every point lands. A model resolves each paragraph neatly before opening the next. A human writer leaves one thread slightly loose, one claim under supported, because that is how thinking actually goes.

• the beats are all the same length
• each paragraph resolves itself completely
• nothing contradicts anything earlier

None of these are proof on their own. A very careful human editor can write evenly. But when three or four line up together on a piece you did not write, the probability is high enough to double check the source.

Here is the move that works best in practice. Read the piece once and ask where a real writer would have disagreed with themselves, added a caveat, or gone off on a short tangent. If there is none, that is the signal.

Keep the document and audience in mind. A technical doc with a fixed template can read even by design. A personal essay that reads that way is far more suspicious. Context is what makes the tell usable.

And do not fixate on one detector. If a tool flags something, run the same sample through two others before you conclude. Disagreement across tools is the strongest hint that the text is human.

To understand why the number is so unreliable, read how to interpret AI detector results. And to pick the right tool for the job, see AI detection tools for writers.

How do you check your own writing for AI tells?

If you're asking how far behind you are, this is easier to answer about a page of text than about yourself. Pick something over 300 words that you wrote this week.

Read it for three things. Do the sentence lengths vary, or do they land in the same range every time? Does any paragraph admit doubt, or leave a point unfinished? Is there one concrete detail, a name, a number, a place, that only you would have put there?

A piece that varies its rhythm, holds something unresolved, and carries a specific detail reads human, even if a tool helped you draft it. A piece that stays even, resolved, and general reads machine-made even when every word is yours. That second case is the common one, and it's fixable in an afternoon.

Detector scores are a poor referee for that second case. A Stanford HAI study that tested seven detectors found 61.22% of TOEFL essays written by non-native English students were flagged as AI-generated, against near-perfect scores on essays by US-born eighth graders.

So keep the three-question read as your first pass, and use a detector only to confirm what you already suspect. A number you can't explain is worse than no number at all.

The false flags are the part most people underestimate, and AI detector false positive rates collects the published numbers side by side.

If you need a free tool to check a sample in the meantime, our roundup of free AI detectors for teachers sets out what each one is actually good at.

One habit worth building now: save a sample of your writing you trust, written before the models got this good. Future you will need a baseline that no machine has touched.

Frequently asked questions