How to Spot AI-Written Content (and Why the Best Tell Is the Easiest to Fake)

How to Spot AI-Written Content (and Why It Sometimes Doesn’t Matter) — The AI Cheat Sheet

The tell everyone teaches you — that AI writing stays vague where a human would get specific — is the easiest one to fake. I found that out by accident. I asked ChatGPT for two paragraphs about a bakery I made up: one in flat marketing voice, one full of quirky specifics. Then I told it that one of the two was written by a human and asked it to pick which. It chose the quirky one, 85% confident. Both were AI. I had just asked for two different voices.

That is the whole problem in one experiment. The tells are real patterns, but every one of them can be requested. So here is what the tells are actually good for, why detector tools are worse than they look, and the short list of situations where any of this is worth your time.

I asked ChatGPT to spot AI writing. It got it backwards.

ChatGPT judging two paragraphs about a bakery and calling one human-written with 85% confidence — Two paragraphs, one verdict. ChatGPT, September 2026. Both paragraphs were written by an AI.

Worth knowing exactly what it got wrong. The reasons it gave for picking paragraph B were the standard ones you will read in every guide to this subject, including the one further down this page: specific sensory detail, an unusual aside, sentence rhythm that varied instead of marching. Those reasons were not stupid. They were just answering a different question than the one I asked — they measure whether writing sounds like a person, not whether a person produced it, and a model will produce that sound on request.

To its credit, ChatGPT flagged its own limits without being asked: it called these “weak signals, not reliable AI detection,” and said an AI told to write like a real customer could easily produce paragraph B. That caveat is the honest part. The 85% is the part people remember.

The tells, and why they are only soft signals

There is no single fingerprint that gives away AI writing, but there are patterns that show up often enough to notice. Think of these as soft signals, not proof.

Everything is smooth and nothing is sharp. AI-generated text tends to be grammatically clean and evenly paced, which sounds like a compliment until you realize good human writing usually has texture. People vary their sentence lengths, break rules on purpose, and let a little personality leak through. When every paragraph is the same tidy shape and every sentence lands with the same rhythm, that flatness can be a hint.

The hedging and the both-sides reflex. AI models are trained to be agreeable and cautious, so they love phrases like it is important to note, there are many factors to consider, and while some argue X, others argue Y. A piece that carefully balances every point without ever committing to one often reads like a machine trying not to offend anyone.

Generic specifics. This one is subtle. AI can produce writing that mentions details but somehow stays vague. A restaurant review might praise the delightful ambiance and attentive service without naming a single dish, a server, or a moment that actually happened. Real experience tends to come with real, slightly random detail: the wobbly table, the waiter who recommended the wrong thing, the song playing when you walked in.

Favorite words and shapes. Certain words and structures come up a lot: delve, tapestry, landscape, navigate, in today's fast-paced world, and the ever-present three-item list. A single one of these means nothing. Several of them stacked together, paragraph after paragraph, starts to feel like a signature.

Openings and closings that say nothing. AI loves to warm up with a throat-clearing introduction and wrap up with a tidy summary that restates what you just read. If you can delete the first and last paragraphs and lose no actual information, a machine may have written them — or a person leaning hard on one.

The honest catch with all of these: humans do every single one of these things too. Tired writers, corporate writers, and people writing in a second language can hit all the same notes. So treat the tells as a reason to look closer, never as a verdict.

When it matters, and when it really doesn't

Here is the part people skip. Not all AI writing deserves the same scrutiny, and getting worked up over the wrong cases just wears you out.

When it matters. School and learning: if the point of an assignment is for a student to think, then handing the thinking to a machine defeats the purpose. This is less about catching cheaters and more about the fact that the learning simply didn't happen. Reviews, testimonials, and recommendations: when you're deciding what to buy, where to eat, or which contractor to trust, you're relying on the idea that a real person had a real experience. Fake or AI-mass-produced reviews poison that well. News and factual claims: AI can state things confidently that are simply wrong, a problem the industry politely calls hallucination. When the content is supposed to inform your decisions about health, money, or the world, you want to know a human checked the facts. Personal or emotional messages: a sympathy note, a wedding toast, an apology. If someone outsourced the feeling to a chatbot, the words might be fine but the gesture is hollow.

When it honestly doesn't matter. A lot of the time, nobody is harmed and nothing is lost — a product description for a phone case, a first draft someone cleaned up with AI help, a routine summary of a meeting, boilerplate instructions for resetting a password. If the writing does its job and the facts are right, whether a person or a tool typed it is beside the point. Plenty of good writers now use AI the way earlier generations used spellcheck and thesauruses: as a helper, not a ghostwriter. Assuming every polished paragraph is a scandal is a fast way to be wrong a lot. The useful question is not was AI involved, it's: does it matter here whether a human was thinking and telling the truth? Sometimes the answer is a firm yes. Often it is a shrug.

Are AI detectors reliable? Why they give false positives

You have probably seen tools that claim to scan text and tell you the percentage that was written by AI. It is tempting to want a clean gauge like that. Unfortunately, as of 2026, these detectors are not trustworthy enough to base real decisions on, and it helps to understand why. Detectors work by looking for statistical patterns typical of machine-generated text, like how predictable each word is given the ones before it. The problem is that this is a guessing game, not a measurement. AI writing has gotten more varied and more human-sounding, and a light human edit can wipe out the patterns a detector looks for. At the same time, plenty of human writing is smooth and predictable, so the tools flag it as machine-made.

That leads to two failures that pull in opposite directions. Detectors miss real AI text, giving false confidence, and they flag real human text, causing false accusations. Neither error is rare. The tools also tend to be biased against certain kinds of writers, including people writing in English as a second language, whose more careful, formulaic phrasing can look machine-like to an algorithm. That is not a small bug; it means the tool can systematically point the finger at the people least able to defend themselves. Even a detector that were 95 percent accurate would still be dangerous when the stakes are high, because a small error rate spread across thousands of essays or reviews means real people get wrongly accused constantly — and there's no way for you, reading the score, to know which cases are the errors. Use these tools, if at all, as one weak signal among many, never as evidence.

If you want evidence for how weak detectors are, the clearest is that OpenAI withdrew its own AI-text classifier in July 2023 for low accuracy, and its guidance for educators says outright that detectors are not reliable enough to act on. The experiment at the top of this page is a small, homemade version of the same failure.

Three ways this goes wrong

False accusations. Do not confront a student, coworker, or writer based on a hunch and a detector score. The cost of being wrong is high — you can damage a reputation, a grade, or a relationship over something you cannot actually prove. If it matters, talk to the person, ask about their process, and look for real evidence rather than a percentage.

Detectors flagging human writing. Remember that clear, well-organized, carefully edited writing is exactly the kind of thing these tools misread. Ironically, the better someone writes, the more suspicious they may look. Punishing good writing because a tool cannot tell the difference is a bad outcome nobody intends but many stumble into.

Over-focusing on detection instead of quality. This is the big one. The interesting question is usually not who or what wrote this, but: is it any good, and is it true? A thoughtful, accurate article helps you whether a human or an AI drafted it. A lazy, wrong one hurts you either way. If you spend all your energy playing detective, you can miss the thing that actually matters, which is whether the content deserves your trust and attention.

What actually works

AI writing is now a normal part of the world, and most of the time that is neither a crisis nor a betrayal. Learn the common tells so your instincts get sharper, but hold them loosely, because humans trip every one of those wires too. Save your real scrutiny for the places where it counts — schoolwork, reviews, news, and anything personal — and let the low-stakes stuff go. Above all, do not outsource your judgment to a detector tool that cannot actually deliver the certainty it advertises. The most reliable instrument you have is still the old one: reading closely, asking whether the words ring true, and noticing when something has a lot of polish but nothing real underneath. That skill works on human writing and machine writing alike, which is exactly why it is the one worth keeping.

Search This Blog