The AI tells: a content guide

This is a guide for non-experts: the giveaways that mark content as machine-made, and what to do instead so your published work still sounds like you.

AI tools are useful for drafting. The problems start when the draft ships unedited, because readers have learnt the patterns and now discount anything that carries them. Below are the tells we currently watch for, with a fix for each one.

Content writing

Contrast statements

“It’s not just a tool. It’s a revolution.”

AI tell. The “it’s not X, it’s Y” construction. AI tools lean on this mirrored pairing so heavily that readers now clock it on sight.

Human alternative. Make the point directly. If a comparison earns its place, spell out what changed and why it matters to the reader. One plain sentence carries more weight than a rhetorical flip.

Opening with a hypothetical question

“Have you ever wondered what your business could achieve with automation?”

AI tell. Leading with a question the reader never asked. It reads as a warm-up lap before the writing starts doing its job.

Human alternative. Open with the answer. Your reader clicked for the substance, so hand it over in the first line. If you’re prompting, ask the tool to lead with the main point instead of building up to it.

Em dashes

AI tell. Yes, we know, we miss them too. The em dash is a perfectly good piece of punctuation that AI models use at several times the human rate, and it has become one of the fastest ways to get a draft flagged.

Human alternative. Swap each one for a full stop or a comma. Two short sentences almost always read better than one long spliced one. You can also tell the tool to avoid them in your prompt, but check the output anyway, because it will sneak them back in.

Lists of three

“Faster, smarter, better.”

AI tell. The rule of three is linguistically smooth, sure, but the robots have worn it out and it no longer reads as genuine.

Human alternative. Use the number of items the point deserves. Two examples can be plenty, and one strong one often beats a set. When a draft hands you a neat triple, cut the weakest item and see whether anyone misses it.

Words that fake authenticity

“We genuinely care about your success.”

AI tell. Words like “genuinely”, “really” and “actually” used as sincerity boosters. This one pains us in particular, because it has long been part of our own human writing style. The robots ruined it for everyone.

Human alternative. Delete the word and reread the sentence. If it still stands, it never needed the crutch. Sincerity comes across through specifics and follow-through, and no adverb can supply it.

Americanised spelling

“Organize”, “color”, “center”.

AI tell. AI models default to American English, so non-local spelling in a Kiwi business’s writing is a giveaway even to readers who couldn’t say why it feels off.

Human alternative. Tell the tool to write in New Zealand English at the start of your prompt, then run your own pass afterwards. The usual escapees are -ize endings and missing u’s.

Emojis, especially in headings

“🚀 Supercharge Your Workflow”

AI tell. A rocket ship in a heading is close to a signature at this point. AI tools sprinkle emojis into business writing far more than any human would.

Human alternative. Strip them from headings without exception. In body copy, use them the way you would in your own messages, which for most business writing means not at all.

Fluff that tries to sound smart

“A rich tapestry of insights about the ever-evolving business landscape.”

AI tell. Ornate phrasing that adds no information. A report is sufficient without being called a rich tapestry of anything.

Human alternative. Name the thing plainly. A useful test: if the phrase would fit any company in any industry, it says nothing about yours. Cut it and state what the reader will learn instead.

“In today’s fast-paced world”

AI tell. We get it, the world is changing quickly. We don’t need to be told at the start of everything vaguely tech related.

Human alternative. Delete the scene-setting opener and start where the substance starts. Most AI drafts begin properly at the second paragraph, so promote it.

Vague enthusiasm without specifics

“A wide range of impressive features.”

AI tell. Enthusiastic adjectives standing in for information. The reader learns nothing about the features or the range.

Human alternative. Say what they are and show the range. List the features, give the numbers, link the examples. Specific claims do the persuading that adjectives can’t.

The robot lexicon

“Let’s delve into how to unlock and elevate your seamless workflows.”

AI tell. “Delve” is the most famous tell in the business. Its use in published writing spiked the moment the chatbots arrived, and it brought friends: “unlock”, “elevate”, “leverage”, “seamless”, “robust”. None of these words are wrong; they’re just fingerprints now.

Human alternative. Keep a banned-words list and run every draft against it. When one turns up, swap it for the word you’d say across a desk. If you’d never say “delve” in a meeting, it doesn’t belong in your newsletter.

The mid-copy question

“The result? More sales.”

AI tell. Different from opening with a hypothetical: this is the punchy question-and-answer fragment dropped partway through to manufacture momentum. The best part? Robots can’t stop doing it.

Human alternative. Join the fragment back into a full sentence and let the fact carry its own weight. If the copy needs a gimmick to stay interesting, the problem is the substance, and no punctuation trick will save it.

Over-formatting

AI tell. A heading every two paragraphs and bullet lists with bold lead-ins on every point. AI output looks like a slide deck wearing a blog post as a costume. Human writing has lumps.

Human alternative. Write in paragraphs by default and promote something to a list only when the reader will scan it, like steps or prices. If every section of a piece looks identical, the structure is doing the talking instead of you.

Throat-clearing transitions

“Moreover, it’s important to note that...”

AI tell. “Moreover”, “furthermore”, “additionally” and “it’s worth noting” standing at the front of paragraphs. Filler connectives occupy the space where an actual link between two ideas should be.

Human alternative. Delete the connective and read the two paragraphs together. If they no longer flow, the fix is reordering the ideas so one leads to the next, and that job belongs to you rather than to an adverb.

The school-essay conclusion

“In conclusion, by following these tips you can...”

AI tell. A closing paragraph that restates everything the reader just finished reading, tied with a bow. Working writers dropped this after high school; the models never left.

Human alternative. End on the last useful thing you have to say. If the piece needs a closer, give the reader an action to take or a consequence to weigh, and cut any sentence that begins with “in conclusion”.

Invented facts and sources

AI tell. The most expensive tell on this page. AI models state statistics and cite sources with total confidence whether or not they exist. A reader who checks one number and finds it invented will discount everything else you publish.

Human alternative. Verify every claim and click every link before the piece ships. If you can’t find the source yourself, the stat doesn’t go in. Better still, feed the tool your own facts and let it do the arranging rather than the research.

Image generation

The ethics sit differently here

AI tell. Image models learned their craft from the work of living artists, most of whom never agreed to the arrangement. The impact lands harder than it does with writing, where training data draws on a far wider and more diverse pool. A generated image can substitute for the exact commission an illustrator or photographer would otherwise have been paid for.

Human alternative. Treat generated imagery as a drafting tool rather than a finished product. When an image matters to your brand, commission a creative and let the AI output serve as the brief. Your budget supports a working artist, and the result belongs to you in a way no prompt can match.

Hands and fingers

AI tell. If you’re using AI to create images, check the number of hands and fingers. We wish we were being silly with needing to list this one.

Human alternative. Count them before you publish. Hands and teeth are the classic failure points. Regenerate or crop anything that looks off, because your audience will spot it faster than you did.

Garbled text

Signage that reads “GRAND OPENIG”.

AI tell. Image models still can’t spell. Any writing inside a generated image, from shopfront signs to product labels, tends to come out warped or misspelt, and readers notice mangled lettering almost immediately.

Human alternative. Keep wording out of your prompts and add any text afterwards in an image editor, where you control the font and the spelling. If a generation arrives with background lettering you didn’t ask for, crop it out or regenerate.

The AI sheen

AI tell. Waxy skin and lighting that flatters everything at once. Generated images share a glossy house style that reads as AI from across the room, especially the corporate ones full of suspiciously attractive colleagues laughing at laptops.

Human alternative. Prompt for the qualities of a real photograph, such as natural light or a specific lens, and reject most of what comes back. Where the subject matters, an honest phone photo taken by a person will outperform a flawless render.

Impossible details

AI tell. Earrings that don’t match and shadows that fall in two directions. The centre of a generated image usually holds up; the edges are where the physics breaks down.

Human alternative. Sweep the whole frame before you publish, paying attention to the background and anywhere two objects meet. Crop tighter if the flaw sits outside your subject, and regenerate if it doesn’t.

Fake photos of real things

AI tell. Generated images posing as photographs of your actual team or premises. When a customer works out that the friendly staff photo never existed, the doubt spreads to everything else you’ve published.

Human alternative. Use real photography for anything that exists in the real world, and keep generated imagery for concepts and illustration, where nobody is being asked to believe it’s a photo. If there’s room for doubt, label it.

Design

The template look

AI tell. Centred hero, gradient text, glassy cards and the same soft purple glow on every AI-built page and ad on the internet. We say this with some discomfort, because we picked our purple long before the robots did.

Human alternative. Start from your brand rather than from the tool’s taste, and change the layout before you change anything else, since structure is where the sameness lives. If a template helps you move fast, treat it as scaffolding and make sure the composition and imagery are yours.

Almost your brand

AI tell. AI tools can’t hold a brand system in their heads. Each asset comes back in a shade near your colours and a typeface near your fonts, and after a month of shipping social tiles and ads your brand has dissolved into an approximation of itself.

Human alternative. Give the tool exact values, meaning hex codes and font names, then check the output against your brand guide anyway. Final assembly belongs in a proper design tool where the values are locked, with AI supplying concepts along the way.

The default fonts

AI tell. Inter and Poppins on every heading, with Playfair Display wheeled out whenever something needs to feel premium. These are fine typefaces that AI tools reach for so reliably that they’ve become part of the look.

Human alternative. Choose type deliberately, the way you chose your logo. Pick a face the templates don’t default to and write it into your brand guide, so every tool you use gets told rather than asked.

The warm filter

AI tell. A yellow overlay laid across every image to make it feel warmer and more human. One golden-hour photo is lovely; a whole site glowing like late afternoon reads as a preset, and the warmth starts to signal the opposite.

Human alternative. Grade images to match your actual brand palette instead of the tool’s idea of cosy. If your photos need a consistent treatment, define one and apply it yourself, and let some pictures just look like the day they were taken.

Cartoon office people

AI tell. Flat vector characters with tiny heads and enormous limbs, high-fiving beside a giant chart. The style predates AI, but the generators adopted it so thoroughly that it now reads as nobody home.

Human alternative. If you need illustration, brief a style that belongs to you, or commission one. Real product shots and real screenshots beat a cartoon of a concept in nearly every ad format.

Icon soup

AI tell. One icon filled and the next outlined, with line weights that wander between them. Generated layouts pull their visual elements from everywhere and nowhere, so nothing quite belongs to the same family.

Human alternative. Pick a single icon library and stay inside it. When a generated design arrives with a mix, swap every icon for its equivalent from your set before anything ships.

Warped type

AI tell. When a model draws lettering instead of setting it, the glyphs come out slightly wrong, with uneven strokes and spacing no typographer chose. It reads as off even to people who couldn’t name a single font.

Human alternative. Never let a model set your type. Generate the artwork without text and add the words in a design tool using real fonts, where kerning behaves and spelling survives.

Almost aligned

AI tell. Elements that sit a few pixels off a shared edge and padding that changes from card to card. Generated layouts approximate a grid without having one, and the eye registers the wobble before the brain can say why.

Human alternative. Rebuild the layout on a real grid before publishing. Use the generation to decide what goes on the page, and use a design tool to decide where.

Spotted one we haven't listed?

This list will keep changing as the tools do. If you’ve spotted a tell we haven’t listed, we’d like to hear about it. Tell us what you’ve noticed.