AI news
AI model reviews
AI tools
How-to guides
Our staffAboutContact

Is this written by AI? Here's the honest answer for 2026

You can't know for sure from a detector score. Here's what AI checkers can and can't tell you, and what actually settles the question.

Person typing on a laptop at a desk
Photo: Christin Hume on Unsplash

You usually can't tell for sure whether something was written by AI, and neither can any tool. AI detectors give you a probability, not proof, and the honest answer to "is this written by AI?" is often "maybe, and here's how confident we can be." If you need certainty, the only reliable routes are a watermark the AI company built in, or the writer's own drafts and history.

That's less satisfying than a green or red light. I think it's also the answer that'll save you from accusing the wrong person.

Short answer

Paste-in checkers estimate how "machine-like" a passage reads. They can be right a lot of the time on long, untouched chatbot output, and they're shaky on short text, edited text and writing by people who learned English as a second language. Treat a score as a reason to look closer; it isn't a verdict.

The background: guessing from style versus checking a signal

There are two very different ways to answer the question, and most confusion comes from mixing them up.

The first is statistical guessing. A language model writes one word (strictly, one token) at a time, picking likely next words. Text produced that way tends to be predictable: low "perplexity", in the jargon, and fairly even sentence rhythm, which researchers call low "burstiness". A detector is a second model trained to spot those patterns. We've got a longer explainer on how AI detectors work if you want the maths, but the key point is simple: it's pattern matching, so a careful human who writes plainly can look "AI-ish", and a chatbot told to vary its style can look human.

The second way is a signal planted at the source. Google DeepMind's SynthID nudges the probability scores of the words Gemini picks so the text carries a watermark you can't see, and Google says it now covers text from the Gemini app and web experience. A matching checker can then look for that hidden pattern. It's closer to checking a banknote's security thread than judging the handwriting (the one analogy I'll allow myself here).

Coverage is the weak spot for watermarks. A SynthID check only tells you about text from models that embed SynthID. Paste in something from a model that doesn't watermark, and the tool has nothing to find.

Real examples

The clearest cautionary tale comes from OpenAI itself. In January 2023 it released a classifier for spotting AI-written text, and on its own announcement it admitted the tool caught just 26% of AI-written text in its test set, while labelling human writing as AI 9% of the time. It also warned the classifier was "very unreliable on short texts (below 1,000 characters)". On July 20, 2023, OpenAI pulled it, citing "its low rate of accuracy." If the company that made ChatGPT couldn't build a dependable detector for its own model, that tells you something.

Turnitin, used by many universities, makes a more confident claim. Its AI writing FAQ sets a target false positive rate under 1% for documents with over 20% AI writing, which it describes as possibly flagging one in every 100 fully human documents. That sounds small. Spread across a course of 300 students, though, it could mean a few people wrongly flagged per assignment. Turnitin also won't score files under 300 words, hides scores between 1% and 19% behind an asterisk, and tells instructors the percentage "should not be used as the sole basis for action". I'd read that last line twice if you're a teacher.

Then there's the vendor marketing. GPTZero's homepage currently advertises "99.9%" accuracy. Maybe that holds on its own benchmark; I'm sceptical it holds on a paragraph a student rewrote three times at 1am. Independent research has been less kind to detectors in general. A 2023 Stanford study led by Weixin Liang, published in the journal Patterns, found that seven popular detectors flagged a majority of TOEFL essays written by non-native English speakers as AI-generated, while doing far better on essays by US eighth-graders. Plainer vocabulary reads as "predictable", and predictable reads as machine.

Here's a small example of my own to show why. Take two sentences:

"The results show that the method is effective and can be used in many situations."
"Weirdly, the method held up best on the messiest data, which nobody on the team expected."

A human could easily have written the first one, especially a tired one or someone writing in their third language. But it's generic and every word is the obvious next word, so a detector would probably score it as more AI-like than the second. The second has a surprise, an opinion and an odd detail. Neither feature proves authorship; they just move the odds.

SignalPoints towards AI?How much it's worth
Very even sentence lengths and stock phrasesSomewhatWeak: lots of formal human writing looks like this
Confident facts that turn out to be wrong or inventedSomewhatModerate, especially fake citations
Detector score over 80% on 500+ wordsYes, probablyModerate, never conclusive on its own
Detector score on under 300 wordsUnclearLow: tools struggle with short text
A positive watermark check (for example SynthID)YesStrong, but only for models that watermark
Version history, notes and drafts from the writerShows the processStrong, in either direction

Benefits and limitations

Detectors aren't useless. On long, unedited chatbot output they seem to do reasonably well, and as a first filter for an editor sorting hundreds of freelance pitches, a score can tell you where to spend your attention. Turnitin's model also tries to flag text that's been run through paraphrasing or "humanizer" tools, which shows the vendors know that's how people dodge them.

The limits are bigger than the marketing admits, in my view. Short text is a coin toss. Mixed text (a human draft polished by AI, or an AI draft heavily rewritten) has no clean right answer, because it's genuinely both. Non-native writers carry extra risk, as the Stanford work showed. And scores move as the models change: Turnitin says it updated its English model architecture in October 2026, so a score from last spring isn't comparable to one today.

So which matters more, catching cheats or protecting honest writers? That's a values call more than a technical one. Turnitin openly chooses to miss some AI text to keep false positives down; it says a document scored at 50% AI could contain as much as 65%. I think that's the right trade-off for schools, where a false accusation can follow a student for years.

Our separate piece on whether AI detectors are accurate goes through the numbers vendor by vendor.

Practical takeaways for readers, teachers and writers

If you're checking someone else's writing, run it through more than one tool and see if they agree. Our roundup of the best AI detectors compares the main options and their limits, and there's a full GPTZero review if that's the one your school uses. Then do the human part: look for invented sources, ask the writer to talk you through an argument, or ask for drafts. A five-minute conversation is probably more revealing than any percentage.

If you're the writer and you've been flagged, don't panic. Keep your drafts and use a tool with version history, like Google Docs, so you can show your process. Ask which tool produced the score and what threshold your institution uses. Point to the vendor's own guidance; Turnitin, for one, says its score shouldn't decide a case by itself.

If you use AI openly (for brainstorming, or to tidy grammar), say so. Many schools and publishers ban passing off AI text as your own work, and disclosure costs you far less than a dispute later. I'd honestly rather read a short note saying "I used ChatGPT to fix my grammar" than spend an afternoon guessing.

Wondering where these signals come from in the first place? Our explainer on how AI is detected in writing covers the tells editors look for by hand.

Next up: watermarks and the EU rules

The bigger shift is coming from the law. Article 50 of the EU AI Act requires providers of generative AI to mark their outputs in a machine-readable way that's detectable as AI-generated, and those transparency duties became enforceable on 2 August 2026. Systems already on the market before that date have until 2 December 2026 to add the marking, and fines can reach EUR 15 million or 3% of worldwide annual turnover. A voluntary EU code of practice sets out how companies can meet the rule, including layered marking of outputs.

That should mean more text arrives with a hidden label, at least from big providers selling into Europe. It won't settle everything. Open-source models can be run without watermarks, and heavy rewriting can weaken a text watermark. My guess is we'll end up with two tiers: text you can verify because it carries a signal, and everything else, where you're back to judgement and probability.

For the second tier, the question "is this written by AI?" may matter less over time than "did the person stand behind it, and is it true?" That's the question I'd start asking now.

FAQ

Can a free AI checker tell me for certain if text is AI-written?

No. Free and paid checkers give a likelihood score. OpenAI withdrew its own classifier in 2023 because it was too inaccurate, and Turnitin tells instructors not to act on its score alone.

How accurate are AI detectors?

It depends on text length, editing and the writer. Vendors claim very high accuracy (GPTZero advertises 99.9%), while Turnitin targets a false positive rate under 1% per document. Independent research has found much higher false positive rates for non-native English writers.

Why was my own writing flagged as AI?

Plain, formal or repetitive writing looks statistically predictable, and predictability is what detectors look for. Short submissions and text you've paraphrased heavily are also more likely to be misread.

Does ChatGPT or Gemini watermark its text?

Google says SynthID now watermarks text from the Gemini app and web experience. EU rules now require providers to mark AI output in a machine-readable way, so I'd expect more providers to follow, though a watermark only helps if a matching checker exists.

What should I do if I'm accused of using AI?

Ask which tool and threshold were used, share your drafts and version history, and offer to explain your work in person. Point out that vendors such as Turnitin say their scores shouldn't be the sole basis for action.

Keep reading