You think you can spot AI writing? Think again. The ghost learned your tricks, stole your style, and vanished before you noticed.
I laugh every time I hear someone brag, "I can spot AI writing in 5 seconds." Really? That's like claiming I can identify every bank robber because he wears a black mask. Criminals change clothes. So do machines. The moment people learn one AI habit, the machines quietly dump it and borrow another. That is the dirty little secret nobody wants to admit. We are chasing yesterday's ghost while today's ghost is already sitting beside us, wearing our jacket, speaking our slang, and signing our name.
For months, the internet swore that em-dashes were the smoking gun. If I saw enough of them, people said, I had found ChatGPT red-handed. Then the ground shifted beneath everyone's feet. Modern large language models no longer flood their writing with em-dashes the way earlier versions often did. The suspect simply changed fingerprints. Case closed? Hardly. The chase has only become more dangerous because many investigators are still searching for clues that no longer matter. They are fighting yesterday's war with yesterday's weapons.
That should surprise nobody. AI is built to learn. Learning means changing. Every software update quietly buries old habits and creates new ones. A detective who memorizes last year's clues becomes tomorrow's fool. You do not catch a chameleon by memorizing yesterday's color.
The evidence keeps piling up. Researchers at The Economist compared human journalism with articles written by OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini, and xAI's Grok. Their study covered 55,940 sentences and about 1.2 million words. That is not somebody's weekend opinion posted on social media. That is a mountain of text large enough to expose patterns invisible to the naked eye.
The results knocked over one popular myth after another.
The real giveaway is no longer endless em-dashes. Instead, today's AI often writes with surprisingly little punctuation. Commas become scarce. Semicolons almost disappear. Parentheses are treated like unwanted relatives at a family reunion. Quotations become rare because AI frequently states information without bringing actual experts into the conversation. The sentences stretch longer and longer, tied together by endless uses of the word "and." Before I reach the period, I sometimes feel as though I have walked across three counties without seeing a stop sign.
That is where many readers get fooled. They expect dramatic clues. Reality delivers quiet ones. AI rarely screams. It whispers.
Vocabulary tells another story. Earlier models loved words like "delve" and endless references to a "rich tapestry." Those phrases became internet jokes. Developers noticed. The models adapted. Now they reach for another bag of tricks. They sprinkle their writing with oversized words such as "significant," "increasingly," "consequences," "interdependence," "parameter," and "methodology." They often convert simple verbs into heavy nouns. Instead of saying people "expanded" something, they prefer "expansion." Instead of saying people "improved," they drift toward "improvement." The sentence grows fatter while the meaning grows thinner.
George Orwell saw this disease decades before artificial intelligence existed. In his famous essay Politics and the English Language, he attacked what he called "pretentious diction." He mocked writers who wrapped simple ideas inside expensive words simply to sound intelligent. Orwell never met ChatGPT, Claude, Gemini, or Grok, yet his criticism lands squarely on much of their output. Machines have inherited one of humanity's oldest bad habits. They often confuse complexity with intelligence.
The irony is almost funny. Humans taught machines to imitate human writing. The machines copied our worst writing first.
Sentence structure exposes another habit. AI likes long paragraphs that keep rolling like freight trains. It dislikes sudden stops. Short, explosive sentences appear less often. Instead, AI reaches repeatedly for familiar rhetorical tricks. It loves "not X but Y." It enjoys "not only...but also." It embraces the classic "rule of three" because grouping ideas into threes usually sounds persuasive. Once I notice those patterns repeating again and again, I stop hearing a human voice. I hear software searching for applause.
Still, I refuse to declare victory that quickly because today's clue may become tomorrow's museum piece.
That is exactly why AI detectors deserve healthy skepticism. Pangram claims 99.98% detection accuracy, an impressive figure on paper. Yet even impressive numbers deserve cross-examination. Detection systems operate largely as black boxes. They often tell me a document is probably AI-generated without clearly explaining why. Worse still, they sometimes accuse innocent writers. False positives are not minor inconveniences. They can damage reputations, punish honest students, and undermine trust. When software becomes judge, jury, and executioner without explaining its verdict, everyone should become nervous.
History offers a warning. Every security technology eventually meets someone determined to defeat it. Spam filters improved. Spammers adapted. Antivirus software improved. Malware evolved. Facial recognition advanced. Criminals experimented with disguises. Artificial intelligence follows exactly the same pattern. Every improvement in detection encourages another improvement in disguise. It is an endless chess match where neither side stays still for long.
The speed of AI evolution makes the contest even more brutal. Only recently, older versions of ChatGPT casually scattered em-dashes across nearly every conversation. Ask today's version the same question, and it recommends using them sparingly. That transformation happened in months, not decades. Human writers usually spend years changing their style. AI changes between software updates. That is like watching an actor leave the stage and return 30 seconds later as an entirely different character.
Many people still imagine AI as a clumsy robot wearing oversized shoes. They expect obvious mistakes. That fantasy grows weaker every month. Large language models learn from enormous collections of human writing and from human feedback. Researchers such as Tommie Juzek of Florida State University note that these systems steadily absorb what readers praise and quietly abandon what readers criticize. Every compliment teaches the machine. Every complaint reshapes it. The ghost studies the people trying to catch it.
That should concern every teacher, editor, publisher, employer, and reader. Looking for one magic clue is wishful thinking. Writing is too complicated for that. Emily Dickinson loved em-dashes long before computers existed. Ernest Hemingway preferred stripped-down prose. Jane Austen had her own fingerprints. Human writers have always broken rules in deeply personal ways. AI now imitates that diversity more convincingly than ever before.
So I no longer waste my time counting punctuation marks like a detective counting cigarette butts at a crime scene. I examine rhythm. I examine sentence variety. I examine whether ideas actually develop instead of merely expanding. I ask whether the writer sounds like someone wrestling with experience or simply rearranging polished language into neat little packages. Machines still struggle to produce genuine insight. They often generate fluent words without earning them.
That difference still matters. For now.
Tomorrow is another story.
The AI ghost is learning faster than most people are reading about it. Every update burns another old clue. Every new model steals another human habit. Every confident expert who says, "I've cracked the code," should remember that ghosts do not survive by standing still. They survive by changing shape before the lights come on. Right now, too many people are staring at yesterday's costume while today's ghost quietly walks through the front door wearing a human face.
As a side note for regular readers, I have also written many titles in my Brief Book Series, now available on Google Play Books. You can also read them here on Google Play, or in Barnes & Noble bookstore: Brief Book Series.
No comments:
Post a Comment