How to Tell If Something Is Written by AI: 12 Signs (2026)

These are the patterns that make readers suspicious. Learn them, because they are a genuinely useful first read. Just hold them loosely, because every single one of them shows up in human writing too.

1. The em dash. The famous one. ChatGPT scattered the em dash, that long horizontal stroke (—), everywhere a comma or a full stop would have done the job. For two years it was the single most cited tell on the internet. Hold that thought, because it is also the clearest example of why these signs do not last.

2. Curly quotes and tidy special characters. Straight quotes turned curly, perfectly formed ellipses, the occasional non-breaking space. Models default to typographically clean output that most people never bother with in a quick email or a message to a friend.

3. The rule of three. Three adjectives, three examples, three bullet points, over and over. Models reach for the triple constantly because the pattern saturates their training data. One set of three is rhetoric. Five sets of three on one page is a habit.

4. The "not this, but that" reversal. Negative parallelism: the construction that denies one thing to affirm another. "It is not just a tool. It is a way of thinking." Used once, it is a clean rhetorical move. Used four times in five paragraphs, it reads like a tic the writer cannot hear.

5. Filler signposting. "It is important to note that." "On the other hand." "In today's fast-paced world." "When it comes to." Phrases that announce a point is coming instead of just making it. They pad the rhythm and say almost nothing.

6. Buzzword vocabulary. Delve, tapestry, testament, realm, harness, underscore, pivotal, illuminate, "navigate the complexities," "the ever-evolving landscape." Any one of these is a normal English word. A cluster of them in a single short piece is a fingerprint.

7. Therapy-speak reassurance. "You are not imagining it." "You are not alone in this." "And honestly?" followed by something that is not especially honest. The chatbot bedside manner, warm and weightless at the same time.

8. Uniform sentence rhythm. Human writing is bursty. A long, winding sentence that builds a thought, then a short one. AI writing tends to hum along at one comfortable length, sentence after sentence. Detectors have a name for this evenness, and it is one of the few things they measure well.

9. Over-structure. A heading for everything, a bullet list for everything, a neat little summary at the end of every section. It reads like a textbook that was afraid to leave a single idea implied or a single connection for the reader to make.

10. The formulaic open and close. "In this article, we will explore." "In conclusion, X remains a complex and multifaceted issue." Bookends with no surprise in them, written to a shape rather than to a point.

11. No lived detail. No specific Tuesday, no named tool, no number that could only come from having actually done the thing. The writing is technically correct and experientially empty. It describes the world from above, never from inside it.

12. Confident fabrication. Citations that do not exist. Quotes nobody said. A well-known fact stated slightly wrong with total assurance. Of every sign on this list, this is the one most worth chasing, because a human can usually explain where a fact came from and a hallucination has no source to point to.

Read that list again and a pattern emerges. Every tell is either something skilled humans do on purpose, or something you can switch off in the model with one sentence of instruction. Often both.

Take the most famous one. Writers love the em dash. It is a real piece of punctuation with a real job, and plenty of careful, published, unmistakably human authors lean on it hard. When ChatGPT made it suspicious, the people most annoyed were the writers who had been using it their whole lives. Then, in November 2025, OpenAI shipped a fix so that ChatGPT and GPT-5.1 finally obey the instruction "stop using em dashes." Sam Altman called it a small but happy win. Think about what that means. The single most cited sign of AI writing is now a setting you toggle off. The tell did not get more reliable. It evaporated on command.

The same is true of the rest. The "not this, but that" reversal is taught in rhetoric classes. The rule of three is older than print; Caesar wrote veni, vidi, vici a long time before anyone trained a language model. Clean grammar and a slightly formal register describe a careful student or a cautious professional just as well as a chatbot. And you can prompt every surface signal away in seconds: tell the model to vary its sentence length, drop the buzzwords, avoid the triple, and write like a tired human, and most of the fingerprint disappears.

There is a darker side to this too. The signs do not fail at random. They fail hardest on the people least able to defend themselves.

A tell you can remove with one sentence of instruction was never proof. It was a habit the model had not yet been asked to break.

Yaps

In 2023, a Stanford team ran 91 essays written by non-native English speakers, real human essays from a standard English proficiency exam, through seven popular AI detectors. The detectors flagged 61% of them as machine-written on average. Nearly all of them, 97%, were flagged by at least one tool, and about a fifth were unanimously flagged by all seven. The likely reason is mechanical: writing by someone still learning the language tends to use a smaller, more predictable set of words, and "predictable" is exactly what these tools read as "artificial." Native English eighth-graders, writing more loosely, sailed through almost untouched.

Sit with that. The clean, careful, slightly plain style that a second-language writer works hard to produce is the style detectors are most likely to call a fake. The tells punish the wrong people.

It helps to know what a detector is really doing, because once you do, its limits stop being surprising.

A detector does not read for meaning. It cannot tell whether an argument is good or an idea is original. It measures two things, and both have plain-English names once you strip the jargon.

The first is predictability. Language models pick words by likelihood: at each step they tend to choose the word that most naturally comes next. So machine text often sits in the safe, average middle of the language, where every word is the expected one. Detectors measure how surprising the writing is, and low surprise nudges the score toward "AI." The trouble is that careful human writing is also unsurprising. A technical manual, a cautious student, a second-language writer, a lawyer trained to be precise: all of them produce exactly the smooth, predictable prose the detector is built to suspect.

The second is variation: how much the sentence lengths and structures change as the piece goes on. Humans vary a lot. Machines, historically, vary less. This one is a slightly better signal than predictability, but it is just as easy to defeat. Ask the model to vary its rhythm, and it does.

So a detector is really asking one question: does this read like the safe, average, predictable middle of written English? For a lot of AI text, the answer is yes. For a lot of human text, the answer is also yes. That overlap is the entire problem, and no amount of model tuning makes it go away.

Now the numbers, because they are worse than the marketing suggests.

No detector is 100% accurate, and the honest range lands somewhere between 65 and 90% on clean, unedited text, falling apart on anything that has been edited or run through a "humanizer." The most damning data point comes from OpenAI itself. The company that builds the models also tried to build a detector for them, and could not. Its classifier correctly caught just 26% of AI-written text while falsely flagging 9% of human writing, and OpenAI quietly shut the tool down in July 2023. If the people who made the engine cannot reliably detect its output, treat every confident third-party claim with deep suspicion.

There is one more thing the vendors do not advertise. An entire category of tools, called humanizers, exists for the sole purpose of rewriting AI text so it slips past detectors. The arms race is real, and it structurally favors evasion, because hiding a signal is easier than proving one is there.

If your competitors handed you a tidy list of detectors with star ratings, here is the same list told straight.

You may have noticed a tool missing from that table. Yaps is not on it, on purpose. Yaps does not detect AI, and a voice app that claimed it could would be exactly the kind of false certainty this guide is warning you about. Yaps sits on the other side of the whole problem, and we will get there. First, the people who feel this most.

If you teach, you are on the sharp end of all of this. A detector flag, a gut feeling, a stack of essays, and a student's record on the line. The pressure to have a clean answer is enormous, and there is no clean answer to be had. So aim for something better than certainty: fairness.

Start with what not to do. Do not accuse on a score alone. You now know that the false positives are not evenly spread. They land on second-language students, on neurodivergent writers, on anyone with a careful, plain, well-organized style. An accusation based on a probability you cannot interrogate, against a student who cannot see the tool that flagged them, is not a process. It is a coin flip with consequences.

Here is what actually works, none of it a piece of software.

Know the writer's voice. The single most reliable signal you have is a sudden jump that a machine cannot account for: vocabulary, structure, or polish that does not match the student's in-class writing, their last draft, or how they talk about the subject out loud. That comparison is something no detector has and you do. Guard it.

Look at the process, not just the product. Drafts, version history, document edit timelines, margin comments, the half-finished outline from two weeks ago. Writing that appears fully formed with no trace of how it was made is the real flag. Writing with a visible, messy history is its own defense, for the honest student and against the dishonest one.

Ask, do not accuse. A two-minute conversation reveals understanding faster than any tool. "Walk me through why you chose this example." "What did you cut from your first draft?" These are easy to answer if the work is yours and very hard to fake if it is not. They also leave a student's dignity intact while you find out.

Design assignments a generator finds hard. Tie the work to a specific class discussion, a personal experience, a local detail, a source you handed out, or an in-progress draft you have already seen. The more a task rewards lived specifics, the less a general-purpose model can help, and the less anyone is tempted to reach for one.

There is a quieter move available too, and it changes the incentive instead of policing it. Encourage students to get their own thinking down by voice. Dictation is not generation. A student who speaks a messy first draft and then shapes it is using their own argument, their own examples, their own voice, just faster than typing it out. It leaves a real process behind, it sounds like them because it is them, and it removes the main excuse students give for outsourcing the writing in the first place: that doing it themselves is too slow. Yaps is built for exactly that, and there are pages for educators and students that go deeper.

Step back from the detection arms race for a second, because it rests on a premise worth questioning.

The entire panic exists because producing text by machine got easy. But text was never the goal. The goal was your thinking, in your words, on the page. There are two honest ways to get there quickly. One is to have a machine generate it for you, which leaves something to detect and a reason to feel uneasy. The other is to say it in your own voice and have it transcribed, which leaves nothing to detect because it is genuinely yours.

The speed argument that pushes people toward generation does not actually hold up. Most people type somewhere between 40 and 80 words a minute. Most people speak at around 150. That is roughly three to four times faster, every time you open your mouth instead of reaching for the keyboard. The honest shortcut was never "let the AI write it." It was "stop typing."

This is where Yaps fits, and it is worth being precise about what it does and does not do.

Be clear about the boundary. Yaps does not write your argument for you, and that is precisely the point. If you want a machine to invent the ideas and shape the prose, that is a different tool with a different ethic, and the detectors will be waiting for whatever it produces. If you want to get your own ideas onto the page faster than your hands can manage, that is dictation, and your words stay yours the entire way.

It runs on Android, macOS, and Windows, transcribes on device by default, and drops clean text into any app you are typing in. There are pages on dictation and a guide for writers if you want the mechanics, and a walkthrough for academic writing if you are doing this for coursework or research.

The honest answer to "was this written by AI?" is that you can have a hunch, not a verdict. Learn the 12 signs and use them as a first read, the way you would notice a stranger acting oddly without convicting them of anything. Distrust any tool, any vendor, any colleague who claims to know for sure, because the technology underneath cannot deliver what they are promising, and the false positives have names and futures attached.

And if what you actually want is good writing produced quickly, notice that the detection problem was always a side effect of one particular shortcut. There is another one. Say it in your own words and let something like Yaps keep up, and the question of whether a machine wrote it answers itself, because the answer is you.

Can you really tell if something was written by AI?

Not with certainty. You can spot patterns that raise suspicion, like heavy em-dash use, buzzwords such as "delve" and "tapestry," uniform sentence rhythm, and a lack of specific lived detail. But every one of those patterns appears in human writing too, and any of them can be removed from AI output with a single instruction. The signs are a reason to look closer, never proof.

What is the most common sign of AI writing?

For a couple of years it was the em dash, the long horizontal stroke that ChatGPT used constantly. It is no longer reliable. In November 2025, OpenAI shipped an update so its models obey the instruction to stop using em dashes, which means the most famous tell is now something a user can switch off in seconds. Plenty of skilled human writers use em dashes heavily, which made it a poor signal even before that.

Are AI detectors accurate?

No detector is 100% accurate. The honest range is roughly 65 to 90% on clean, unedited text, and accuracy drops sharply on anything edited or run through a "humanizer." OpenAI's own detector caught only 26% of AI text while falsely flagging 9% of human writing, and the company shut it down in 2023. Treat any vendor claim of near-perfect accuracy with deep suspicion.

Why do AI detectors flag human writing as AI?

Because they measure predictability, not authorship, and careful human writing is often very predictable. A 2023 Stanford study found that popular detectors falsely flagged about 61% of essays written by non-native English speakers as AI, because writing in a still-developing second language tends to use a smaller, more predictable vocabulary. Technical writers, cautious students, and precise professionals get caught for the same reason.

Is the em dash still a sign of ChatGPT?

Not reliably. It was a strong tell through 2024 and into 2025, but two things broke it. First, users learned to prompt it away, and as of late 2025 OpenAI's models obey that instruction directly. Second, the backlash reminded everyone that many human writers love the em dash and always have. A page full of them might be a machine, or might be a careful writer. It tells you nothing on its own.

Can teachers tell if a student used ChatGPT?

They can suspect it, not prove it from the text alone. The strongest tools a teacher has are not software: knowing a student's usual voice, noticing a sudden unexplained jump in quality, looking at the draft and version history, and asking the student to walk through their own argument. A detector score should never be the basis for an accusation by itself, given how often those scores are wrong about the most vulnerable students.

What words are dead giveaways of AI writing?

Common offenders include "delve," "tapestry," "testament," "realm," "harness," "underscore," "pivotal," and "illuminate," along with phrases like "navigate the complexities" and "the ever-evolving landscape." A cluster of them in a short piece is suspicious. But these are ordinary English words, and a single instruction tells the model to avoid all of them, so their absence proves nothing either.

Can AI-written text be made undetectable?

Effectively, yes. A specific category of tools called "humanizers" exists to rewrite AI text so it slips past detectors, and careful prompting alone removes most surface tells. This is the core reason the detection arms race favors evasion: hiding a signal is structurally easier than proving one is present. It is also why no detector can promise reliable results against someone who is actively trying to avoid it.

Will I get in trouble for a false positive on an AI detector?

It happens, and it is one of the strongest arguments against trusting these tools. To protect yourself, keep evidence of your process: drafts, version history in your document editor, notes, and outlines that show the work taking shape over time. If you dictated rather than typed, that capture history is part of the trail too. A visible, messy process is the best defense against a false flag.

How can I prove I wrote something myself?

Show the process, not just the product. Keep dated drafts, leave your document's version history intact, save your outlines and research notes, and be ready to explain your choices out loud. Writing that has a visible, uneven history of edits is far more convincing than a single finished file, because that history is exactly what a one-shot generation does not produce.

Does Google penalize AI-generated content?

Not for being AI as such. Google's stated position is that it rewards helpful, reliable, people-first content and demotes low-value content regardless of how it was produced. In practice that means thin, generic, derivative writing tends to lose, whether a human or a machine made it, while genuinely useful content can rank either way. "Was it AI" is the wrong question; "is it actually useful and trustworthy" is the one that matters.

Is it cheating to use dictation instead of typing?

No. Dictation transcribes the words you say into text. The ideas, the argument, the phrasing, and the voice are all yours, exactly as they would be if you typed them, just produced faster. That is fundamentally different from asking a model to generate the content for you. Speaking your own draft is a writing method, not a shortcut around writing.

Does using Yaps count as AI writing?

No. Yaps turns your speech into text. It does not invent ideas, compose arguments, or write paragraphs you did not say. The output is your own words, captured by voice instead of keyboard, which is why dictated drafts carry the natural rhythm and specific detail that AI-generated text lacks. There is nothing for a detector to find, because the writing is genuinely yours.

What is the best AI detector for teachers?

The honest answer is that no detector should be used alone, so the "best" one is whichever you treat as a single weak signal rather than a verdict. Tools like Turnitin and Pangram advertise lower false-positive rates than free options like GPTZero, but all of them make mistakes, and all of them misjudge non-native and neurodivergent writers more often. Pair any tool with knowledge of the student's voice, a look at their drafting process, and a direct conversation before drawing any conclusion.