How to tell if a text was written by AI
By Chatday Editorial Team ·
Here’s a number worth sitting with: researchers who sampled tens of thousands of web pages found that roughly half of all new articles published online are now written by AI. Not a fringe corner of the internet. Half. Which means you’ve almost certainly read AI text today, probably without noticing, in a product review, a news roundup, a LinkedIn post, maybe an email.
So how do you actually tell? The honest answer is that you can get quite good at it, with a mix of your own eyes and the right tool, as long as you know what each one can and can’t prove. This guide gives you both: the signs a human can spot in thirty seconds, and how AI detectors work when you want a second opinion.
Half the internet is now written by machines
The stat in the intro comes from Graphite, an SEO research firm that analyzed tens of thousands of English-language articles published since 2020, running each one through several AI detectors. Their finding: the share of AI-generated articles exploded after ChatGPT launched in late 2022, then leveled off at roughly 50% through 2025 and into 2026. Some quarters it edges just above half, some quarters just below.
And it’s not only blog spam. A separate analysis of more than 15 million biomedical paper abstracts found that by 2024, at least 13.5% of them showed signs of being written or polished by AI, based on telltale vocabulary that spiked after ChatGPT arrived. If it’s happening in scientific journals, it’s happening everywhere.
None of this means AI text is automatically bad. Plenty of it is useful, and plenty of humans use AI as a drafting partner and then edit heavily. But when you’re grading an essay, reading a glowing product review, or deciding whether a news article is trustworthy, “who actually wrote this?” is a fair question. Here’s how to answer it.
The signs you can spot yourself
Before any tool, use your eyes. AI models learned to write from an enormous pile of internet text, then got polished toward a tone that raters liked: smooth, agreeable, slightly formal. That polish became a fingerprint. Here’s what to look for when you’re reading someone else’s text.
| The sign | What it looks like |
|---|---|
| Stock phrases | ”In today’s digital landscape”, “it’s important to note”, “let’s dive in”, “game-changer” |
| Suspiciously even rhythm | Every paragraph the same size, every sentence the same medium length, tidy lists of three |
| Confident but empty | Lots of polished sentences that could apply to any topic, with no lived detail |
| No verifiable specifics | No names, dates, prices or first-hand moments, or ones that turn out to be wrong |
| Perfect grammar, no voice | Zero typos, zero contractions, zero personality, like a press release about nothing |
| The formula addiction | ”It’s not just X, it’s Y”, “from X to Y”, a rhetorical question as the opener |
The single strongest human check is the specifics test. Real writers, even mediocre ones, leave fingerprints of actual experience: a specific store, a weird detail, an opinion that costs something. AI defaults to text that sounds right rather than text that is right. Pick one or two concrete claims and check them. If a “review” of a restaurant never mentions a dish, or an article cites a study you can’t find, your suspicion is earned. That failure mode has a name, and we’ve covered why AI makes things up in plain language before.
One caveat before you start accusing people: every sign in that table is a habit, not a smoking gun. Humans wrote in stock phrases long before ChatGPT did. Which is exactly why a second opinion helps.
How AI detectors actually work
An AI detector reads a text and asks one core question: how predictable is this?
Language models write by choosing, again and again, the most statistically likely next word. That leaves a trace. AI text tends to be smooth and predictable, with few surprising word choices and an even, steady rhythm. Human writing is messier. We start sentences over, pick odd words, go long and then suddenly short. Detectors measure that predictability (researchers call it “perplexity”) and the variation in rhythm (they call that “burstiness”), and produce a probability: this text looks 85% likely to be AI-generated.
The key word is probability. A detector never actually knows who typed the words. It’s making an educated statistical guess, the same way you were in the section above, just with math instead of vibes.
That’s why the score behaves the way it does:
- Longer texts give better readings. A two-sentence message simply doesn’t contain enough signal. Most detectors are unreliable below a few hundred words.
- Mixed texts confuse it. A human draft polished by AI, or an AI draft edited by a human, lands in the grey zone, because it genuinely is both.
- Formal writing scores “more AI”. The more predictable and conventional the style, the more machine-like it reads, even when a person wrote it.
Where detectors get it wrong
This is the part most articles skip, and it’s the part you most need. AI detectors have real, documented failure modes.
The most famous example comes from OpenAI itself. In early 2023 the company launched its own AI text classifier, and shut it down six months later citing its “low rate of accuracy”. By its own numbers, it caught only 26% of AI-written text and wrongly flagged human writing as AI 9% of the time. If the company that built ChatGPT couldn’t reliably detect ChatGPT, treat any tool promising certainty with suspicion.
The second failure mode is worse because it’s unfair. A Stanford study published in the journal Patterns tested seven popular AI detectors on essays written by real people. For essays by native English speakers, the detectors did fine. For essays written by non-native English speakers, the detectors flagged them as AI-generated 61% of the time. The reason is built into the method: people writing in a second language tend to use safer, more conventional phrasing, which reads as “predictable” to the math. Careful, formal, by-the-book writing is exactly what detectors think a machine sounds like.
How to check a text, step by step
Put together, here’s the routine that actually works. It takes about two minutes.
- Read it once for the tells. Stock phrases, even rhythm, formula openers, no personality. Note your gut feeling, but don’t stop there.
- Run the specifics test. Find one or two concrete, checkable claims. Search for them. Invented sources, wrong prices and unverifiable studies are the strongest red flag there is.
- Run it through a detector. Paste the text into an AI detector and get the probability score. Free, takes ten seconds, no signup needed to try it.
- Weigh the length. Under a couple hundred words, discount the score heavily. On a long, unedited text, give it real weight.
- Consider who wrote it. Formal register or a non-native writer? Shift your threshold accordingly, because that’s exactly where false flags cluster.
- Combine everything. Human tells + failed specifics + high score = very probably AI. One signal alone = keep your accusation to yourself.
When it matters, and when it honestly doesn’t
Not every AI-written text is a problem. A weather summary or a product description written by AI harms nobody. The question matters in specific situations:
- Teachers and professors checking whether work is a student’s own. Use the routine above, and remember the Stanford finding before acting on a score alone.
- Hiring managers reading suspiciously polished cover letters. A detector plus one specific interview question (“tell me more about this project”) settles it fast.
- Buyers reading reviews. Five-star reviews with zero product specifics are the classic pattern.
- Anyone reading news. If an article’s facts don’t check out, whether a human or a machine wrote the mistake matters less than the mistake itself.
And there’s the flip side. If you write carefully in a second language, or just have a formal style, your honest work may get flagged one day. It helps to know your own baseline: run your writing through the detector and see how it scores. If your genuinely human text keeps reading as robotic, that’s a style problem more than a fairness problem, and we’ve written a whole guide on how to make AI writing sound human whose advice works just as well for humans who write like machines. The AI humanizer can rework a stiff draft in one click.
The two-minute habit worth building
Half of what you read is now machine-made, and that share isn’t going back down. The good news is that you don’t need to become a forensic linguist. Read for the tells, check one fact, and let a detector do the statistical part. Together, those three steps catch far more than any single one, and they keep you from the one real mistake in this game: being confidently wrong about a real person’s writing.
The next time a text feels a little too smooth, don’t just wonder. Check.