Spot the AI Text

Why AI text is everywhere now

A few years ago, generating a paragraph of fluent, on-topic prose took a human being. Today it takes a sentence of instruction and a couple of seconds. The result is that AI-written text now sits quietly inside product descriptions, news summaries, cover letters, student essays, marketing emails, customer reviews, and a large share of the fresh content published to the open web every day. Much of it is harmless. Some of it is not: fake reviews, fabricated sources, and plausible-sounding misinformation all get cheaper to mass-produce when the writing is free.

Naturally, people want a way to tell the difference — a tell, a tool, a tick-box that says “a machine wrote this.” The uncomfortable truth, which the game above is designed to teach by experience, is that reliable detection by eye is mostly a comforting myth. You can raise your suspicion, but you cannot manufacture certainty. Understanding why is more useful than any single trick.

The tells of AI writing

AI writing does have recognisable habits, especially in its default, un-edited state. The most common ones:

• Over-hedging.Phrases like “it’s important to note,” “that said,” “while there are many factors” and “there is no one-size-fits-all solution.” The model is trained to avoid committing, so it qualifies everything into blandness.
• Suspicious symmetry and tricolons.Neat lists of three (“streamlining processes, enhancing efficiency, and delivering value”) and balanced sentences that feel almost too tidy. Real writing is lumpier.
• Generic specifics. Text that sounds concrete but names no actual thing: “across various industries,” “a leading provider,” “in today’s digital landscape.” Nothing you could look up or check.
• Uniform rhythm.Sentence after sentence of similar length and shape. Human writers vary wildly — a long winding clause, then a short one.
• Signature vocabulary.An over-fondness for “delve,” “tapestry,” “underscore,” “leverage,” “foster” and long em-dashes used as connective tissue.
• No lived detail.The biggest one. Human writing tends to carry a specific, slightly pointless, checkable fact — a five-cent washer, a train missed by ninety seconds, a flat clarinet section. AI text is fluent but strangely weightless, because it never actually did anything.

Now the crucial caveat: any single one of these tells is unreliable. Plenty of humans hedge, love a tricolon, and adore the em-dash. Plenty of AI text has been edited to strip the tells out entirely, or was prompted to write with grit and specifics. Treat the list as a way to raise a question, never to close one. The moment you feel certain from surface signals alone is exactly the moment you are most likely to be wrong.

Why AI detectors don’t work

If your eye is unreliable, surely software can do better? It cannot, and the reasons are structural rather than fixable with a better model.

• High false-positive rates. Detectors routinely flag genuine human writing as AI. When the cost of a false accusation is a failed assignment or a rejected application, even a small error rate is unacceptable.
• Built-in bias.Detectors tend to score simpler, more formulaic English as “AI-like,” which systematically penalises non-native English writers and some neurodivergent writing styles. They punish people for how they naturally write.
• Trivially defeated.Light editing, a paraphrasing pass, or asking the model to “write less like an AI” is usually enough to drop a detector score to zero. The people trying to hide AI use are the ones a detector fails to catch.
• No ground truth. There is no signature, watermark or fact of the matter that a detector can measure. It is guessing from style, and dressing that guess up as a confident percentage.

The practical advice is blunt: do not trust AI-detector scores for any consequential decision— grading, hiring, moderation, publishing. A number that looks precise is still just a hunch, and often a biased one.

What actually works

If detection is a dead end, what should you do instead? Shift the question. Stop asking “did a human write this?” and start asking “is this true, and can I check it?” That question has real answers.

• Verify the claims and sources. Follow citations to their origin. AI text frequently invents sources that look real and evaporate the moment you search for them.
• Look for specific, checkable facts. Dates, named people, numbers, quotes you can confirm. Vagueness is a red flag whoever wrote it.
• Consider context and provenance. Who published this, where, and why? A brand-new site with a flood of generic articles behaves differently from a named author with a track record.
• Ask the author. In a classroom or a workplace, a conversation about the work reveals far more than any tool. Someone who understands what they submitted can talk about it.

And remember that text is only half the problem. AI-generated images and deepfake video raise the same questions with a different set of tells — hands, teeth, lighting, reflections and metadata. If that is your worry, work through our companion checklist for how to spot AI images & deepfakes. The underlying discipline is the same in both: verify, don’t just vibe-check. Train your eye with the game above, but let the verifying — not the guessing — be what you actually rely on.