How to Tell If a Text Was Written by a Human or AI? The Economist Breaks Down Three Key Flaws to Spot 'AI Ghostwriting'

An all-pervasive 'linguistic ghost' is spreading across publications and the internet—LinkedIn profiles, academic theses awaiting oral defense, and even entries in competitive literary awards may all bear the invisible hand of generative AI. But the real question remains: do we have reliable ways to instantly tell whether a given text was written by a human or generated by a large language model (LLM)?

According to The Economist, AI is rapidly infiltrating the English-speaking world as a 'ghostwriter.' It can handle essays, poetry, news reports, and corporate jargon with ease. It can mimic Shakespearean sonnets or the tone of pulp fiction; it can replicate Hemingway’s concise prose or draft an office printer manual without breaking a sweat. Its writing speed is astonishing—professional writers have reason to fear, as even Hemingway rarely produced such volume in a single day, and he needed far more alcohol to do so. Statistics show that over one-third of new website content is now generated by artificial intelligence.

So how can we detect AI ghostwriting? The Economist notes several methods, including using trained detection algorithms. Pangram, a leader in this field, claims its accuracy reaches 99.98%. However, such detection tools operate as black-box algorithms, risking false positives, and companies often fail to explain how conclusions are reached. Other researchers attempt to search for suspicious vocabulary or compare AI-edited versions of texts, but they struggle to isolate the true stylistic fingerprints of AI writing.

Compared to these approaches, The Economist clearly favors analyzing differences in writing style between humans and machines. The publication conducted a large-scale experiment, comparing 1.2 million words and 55,000 sentences generated by mainstream AI models like ChatGPT, Claude, Gemini, and Grok, against human-written texts from The New York Times, The Washington Post, CNN, and bestselling novels published between 1950 and 2022.

The study found that AI-generated text differs significantly from human writing in vocabulary, punctuation, sentence structure, and paragraph organization. However, this doesn’t mean AI has a unique or superior writing style. The Economist argues that AI writing often lacks clarity and elegance, tends to be monotonous, and aspiring writers should avoid these AI-style flaws. Unfortunately, these flaws aren’t permanent—AI evolves at an astonishing pace with each model update or prompt adjustment.

Nonetheless, The Economist summarizes three key flaws to identify AI writing:

**Flaw One: Preference for 'Pretentious Diction'**

While older AI models frequently used words like 'delve,' newer versions have reduced such markers. Instead, they now heavily favor multisyllabic words, obscure terms, and scientific jargon—such as 'significant,' 'increasingly,' and 'consequences.' They also use more rare words like 'interdependence' and 'reindustrialisation,' and technical terms like 'parameter' and 'methodology.'

AI also favors nominalization—turning verbs into nouns (e.g., changing 'expand' to 'expansion').

George Orwell, author of 'Animal Farm' and '1984,' once criticized 'pretentious diction'—the use of complex vocabulary to dress up simple statements. He noted that such writers often believe Latin or Greek-derived words are inherently more sophisticated than Saxon-origin words. Interestingly, AI seems to agree: AI-generated texts use Latin-derived vocabulary at a much higher frequency than human writing.

**Flaw Two: Reduced Dash Use, But Overall Punctuation Avoidance**

The popular belief that long dashes are a hallmark of AI writing is outdated. The Economist’s empirical analysis shows that, except for Claude, ChatGPT actually uses dashes less than humans. However, a new flaw emerges: AI drastically underuses punctuation overall.

Because AI tends to write long sentences and relies heavily on 'and' as a connector, it uses far fewer commas, semicolons, and parentheses than human writers. Additionally, since AI rarely quotes experts directly, quotation marks appear less frequently.

**Flaw Three: Lack of Short-Sentence Impact and Overreliance on Writing Formulas**

Skilled speakers and writers excel at alternating long and short sentences, using brief, powerful statements to energize the rhythm. But The Economist notes that AI-generated paragraphs are often composed of long, flat sentences. When AI tries to make writing more dynamic, it heavily relies on fixed rhetorical formulas—such as 'not X but Y,' 'not only... but also,' and the 'rule of three.'

The Economist advises that, for now, one should watch for writing that is dull, pretentious, and filled with Latinate vocabulary. However, it also warns that with each update, AI writing is becoming increasingly indistinguishable from human writing.

Tommie Juzek from Florida State University’s linguistics department points out that large language models (LLMs) are being trained to learn from human usage and feedback, absorbing what people find compelling and discarding weaker content.

The Economist gives an example of AI evolution: asking an older model, 'Does AI overuse dashes?' would only elicit a vague 'Good question!'; today, the same query receives a serious response: 'Dashes should be used sparingly.' The magazine quips that this pace of evolution isn’t human—it’s ghostly, fitting for a 'ghostwriter.'