Most people meet machine-written text the same way they meet spam. They assume they would recognize it on sight.
The evidence says the opposite. Pew Research Center ran almost half a million archived English-language pages through a detection model and found that about one in ten showed significant signs of AI authorship as of July 2026. Filter the sample to pages published after ChatGPT launched and the share rises past one third.
You read this material every day. Product reviews, recipe preambles, local news sites, LinkedIn thought posts, the About page of a business you were about to hire. Some of it is harmless. Much of it exists to win a spot in search results, and an unknown slice is built to push a scam or a fake rating. The skill that matters now is knowing when your guess is unreliable, learning to notice the signs that actually carry weight, and checking a text the way an editor would instead of trusting a first impression.
Your instincts on this are measurably weak
Put a large group of ordinary people in front of human and AI writing and they land near a coin flip. A study published in PNAS tested 4,600 participants rating self-presentations written for job, hospitality and dating contexts. Accuracy sat between 50 and 52 percent. Paying people for correct answers did not help. Giving them instant feedback on each guess did not help.
The study also showed why. People leaned on cues like first-person pronouns, contractions and everyday topics as proof of a human author. Those signals carry no weight. Machines reproduce them easily, and text tuned to those assumptions gets judged as more human than human. Experiments with short stories found readers performed at or below chance at picking the human author, and reading more carefully made them worse, not better.
One finding points somewhere useful. People who regularly generate and edit text with these tools get better at recognizing the pattern, and familiarity with a subject helps. This is a learnable skill. It just is not one you pick up from a vague feeling that a paragraph seems off.
Large language models write by predicting the most probable next word, over and over. That mechanical habit produces a house style. When a model faces a choice between a plain word and a polished one, it reaches for the polished option every time, because the polished option carries higher probability in its training data. A person would find that repetition grating and would vary it. The model cannot feel the repetition.
The effect shows up in hard numbers. A team led by Dmitry Kobak tracked vocabulary in more than 15 million biomedical abstracts from 2010 to 2024 in a study published in Science Advances. Words like delve, underscores and showcasing appeared far more often after ChatGPT arrived than the pre-2023 trend predicted, with the delve word family roughly 28 times more frequent. The same analysis estimated that at least 13.5 percent of 2024 abstracts had been processed by a large language model, a lower bound that reached 40 percent in some subfields. The vocabulary shift was larger than the one the pandemic caused.
None of this means a single word proves anything. People used these words before 2023 and still do. The signal is density, plus what sits around it. A short piece that stacks three or four of these markers inside otherwise flawless prose is far more suspicious than one that uses a single favorite word.
- Filler verbs for ordinary actions: delve, underscore, showcase, highlight, leverage, foster
- Grand nouns for plain things: tapestry, realm, landscape, testament, cornerstone, journey
- Formal intensifiers: pivotal, crucial, meticulous, intricate, multifaceted
- Polite signposts: it is important to note, it is worth mentioning, in conclusion, overall
Signals that survive a normal read
Vocabulary is the easiest tell to fake, because writers have learned to avoid those words. Structure is harder to change. The table below covers what to look for and why it shows up. Judge the whole set, never a single row.
| Tell | What it looks like | Why it happens |
|---|---|---|
| Over-clean rhythm | Paragraphs of near equal length, sentences all the same size | Prediction favors steady medium-length sentences. Human writing swings between a one-liner and a long sprawl. |
| Punctuation by habit | Em dashes where a comma or a full stop would do, Oxford commas in every list | Pew measured em dashes appearing about twice as often on the web since 2023. AI models overlearn these flourishes. |
| Announced structure | Every section ends with a tidy summary that restates the opening | The model was trained on essays and articles that modeled a proper wrap-up, so it closes even when nothing needs closing. |
| Nobody is in it | No names, dates, prices, places or odd small details | The model has no memory of a life. Unprompted specifics take effort and risk error, so they rarely appear. |
| Calm symmetry | Each claim meets a balanced counterclaim before a neat landing | Training rewards helpful neutrality. Taking a side looks risky to the model. |
| Balanced lists of three | Arguments and examples arrive in tidy trios | Models favor parallel structure, and trios read as complete to the probability engine. |
Word-frequency claims draw on Kobak and colleagues, Science Advances, July 2025. Punctuation trends draw on Pew Research Center, August 2026. Figures checked in September 2026.
Detection tools help, until you treat them as a verdict
Automated detectors work on probability, not certainty. They compare a text against statistical patterns in how humans and machines choose words, then return a score. OpenAI retired its own classifier in July 2023 after about six months because of low accuracy. At launch it caught roughly 26 percent of AI-written text and wrongly flagged about 9 percent of human text. Independent evaluations since then keep finding the same shape of problem. Lightly edited human writing gets flagged at high rates, while AI text passed through a rewriting tool slips past detection in the large majority of cases. A writer determined to hide the origin usually can.
The tools also disagree with each other. Pew noted that its two detection models gave different answers on many individual pages even though the overall trend matched. A single percentage is a weak basis for a judgment about one specific person.
They are still worth using as a second opinion. When you want a fast read, an AI detector free to run, like the one at ZeroGPT, returns a percentage in seconds. Treat the number the way you would a weather forecast. It gives direction and probability, not certainty. If two tools agree, read the text again with fresh suspicion. Do not treat agreement as proof.
False positives carry real cost, which is why a score should never end a conversation. In 2023 a professor at Texas A&M University-Commerce fed student essays to ChatGPT itself and concluded the chatbot had written all of them, even after it produced the same yes-I-wrote-this answer for a passage from the professor’s own dissertation. The students were cleared. A University of North Georgia student whose paper was flagged after she used Grammarly’s grammar checking reported losing scholarship eligibility and being put on academic probation. Research also finds that polished writing by people who learned English as a second language draws more false flags than native writing. The tools measure style. Style is not the same as authorship.
A reading routine that takes a few minutes
You do not need to become an expert. You need a repeatable way to check text that matters to you.
- Start with the substance, not the style. Skim for verifiable specifics. Names, dates, prices, places and numbers are things a writer must get right, and getting them right is effort that automated filler rarely spends. A paragraph of smooth generalities with nothing you could check is more suspicious than any single word.
- Count the markers, then the gaps. Three or more vocabulary tells in a short piece raises the odds. Then look for human residue. An awkward phrase, an aside, an opinion defended with a reason, a detail only a participant would know. Machine text reads fluent and empty. Human text is lumpy and specific.
- Watch the rhythm. If every paragraph runs about the same length and every point meets a tidy counterpoint before a summary line, someone tidied the thinking. Real writing lurches.
- Quote an odd sentence into a search engine. Content farms reuse text and prompts recycle phrasing. A distinctive line that appears verbatim across unrelated sites is a strong signal.
- When the stakes are real, run a detector as one input. A score on its own should not fail a student, fire a freelancer or decide a purchase. It should send you back to the text for a closer look.
Machine writing has crossed the line where authorship alone tells you much. Useful AI text gets published daily and reads fine. Human text fails daily too, full of confident error with no machine involved. The question with teeth is what the writing does for you. Is it specific enough to check, are its claims verifiable, and does the writer have any reason to be careful? Ask that instead of who typed this, and you will be harder to fool than anyone who memorized a word list. The tells change every few months as models improve. The empty, unfalsifiable paragraph does not.