Four signs a text was written by AI, and five that a website was

The detectors are close to guessing, so the tells have to be read by eye - and it is the same exercise whether the machine wrote the words or drew the page

There is no reliable way to prove a particular text came out of a chatbot.

Start by not trusting the detectors

The services that sell this as a product are close to guessing¹. They will cheerfully tell you the Bible is 88% ChatGPT. Worse, they are not wrong at random: a study of seven detectors found they flag writing by non-native English speakers as machine-written far more often than the same content by native speakers². OpenAI withdrew its own classifier for exactly this reason - low accuracy, admitted publicly³.

So the signs below are not proof. They are what a reader can notice without any service at all.

Four signs in a text

1. The words themselves. ChatGPT likes to open with "in today's world" and cannot leave "an integral part of" alone. The vocabulary markers are familiar by now: delve, streamline, fostering, realm, ever-evolving, robust. This is measurable rather than anecdotal - a study of fourteen million biomedical abstracts found the frequency of exactly this set of words jumped after 2022, at a rate previously seen only during the pandemic⁴. Why these words? They were probably frequent in the "successful" professional prose the model learned from: high-register clichés with conveniently vague meanings.

2. Spelling and paragraphs that are too tidy. A living author drops a capital, forgets a closing quotation mark, leaves one bullet without a full stop. The machine never does, and it will always cut the text into even, print-ready blocks. Students writing dissertations trade advice on Reddit about how to scatter small untidinesses through a text so the committee does not read it as ChatGPT.

3. A flat, slightly wooden register. A person's style jumps about: slang, jargon, whatever phrase is fashionable this month. By default a language model writes like a well-behaved term paper. Careful instructions fix this, but not everybody bothers.

There is a subtler version of the same thing. The machine has no ego and no life, so it does not reach for "as far as I can tell" or "in my experience."

4. The watery paragraph at the end. The model's job is to be of service, so it closes by summarising what it just said, in case you lost the thread. People stop where their thought stops. When you paste an answer in a hurry, check whether that final paragraph came with it - it is the loudest tell of the four.

Wikipedia maintains a far longer catalogue of these, written by editors who spend their days reverting machine-written articles⁵.

Five signs in a website

Nothing is wrong with building a site with AI in 2026. The question is whether it was done in earnest or in fifteen minutes. A landing page generated by a beginner is as easy to spot as an em dash in a ChatGPT paragraph.

1. The vibecode palette: purple on a dark background. Colour gives it away first - purple, magenta, cyan. Call it tech-decor. The model probably learned it from Dribbble, off the designer portfolios of the early 2020s, when purple was in fashion.

2. Polar gradients and gradient blobs. My guess is this came out of Apple's brand guidelines. It is the machine's attempt to make your site look expensive. Human designers use a careful gradient to signal corporate premium; it is no surprise the model cosplays the trick.

3. The Claude aesthetic: a beige site in a craft serif. Instrument Serif, Fraunces. Anthropic's model reaches for the opposite of cyberpunk neon, as if trying to hide the machine behind something handmade: a noble cream palette, soft rounded lines, vintage humanist faces from the free end of Google Fonts.

4. Pills and fake charts. A pill is the capsule-shaped interface element the model wants to wrap around everything. It loves cards too, and overuses them - cards inside cards inside cards. Then there are the sparklines: little charts climbing up and to the right, plausible at a glance, carrying no information whatsoever.

5. A wall of text and no pictures. I asked a product designer in Estonia how he recognises them⁶.

Besides the purple gradients and the fake premium look, an AI landing page usually has a great deal of text and no images or diagrams. Simply because AI services cannot generate those assets by default.

What actually settles it

None of this is evidence, and it will age: the next model will have different habits, and the beige serif will look as dated as the purple gradient does now. What does not age is provenance - a signature travelling with the file rather than a guess made about it⁷. That is the machinery the European Union has just made compulsory, which is a better answer than any detector⁸.

Until it is everywhere, this is what is left: read carefully, and look at the last paragraph.