How to Detect AI-Generated Writing
The Economist's research reveals distinctive features of AI writing.
1Research Scope
The Economist conducted a comprehensive study comparing human and machine writing. Leading language models including OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini, and xAI's Grok were asked to write articles. The study analyzed 55,940 sentences and 1.2 million words. AI texts were also compared with journalism samples from CNN, New York Times, and Washington Post.
2Methodology
In the research, AI models were given summaries of Economist articles and asked to write original pieces. This created a corpus of both human and AI-generated texts. AI writings were also tested against excerpts from bestselling novels published between 1950-2022. This multifaceted approach clearly revealed differences in AI writing style.
3Word Choice Differences
AI writing tends to use more Latin-derived words compared to human authors. The 'pretentious diction' that George Orwell described - dressing simple statements with complex words - is prominent in AI texts. AI also prefers Latin or Greek-derived words over Saxon ones. This creates significant artificiality in writing style.
4Punctuation Usage
Contrary to popular belief, after recent updates AI models no longer overuse em-dashes. Only Claude uses more em-dashes than human writers. ChatGPT uses significantly fewer than any other writer. AI texts generally contain less punctuation - particularly fewer commas and semicolons compared to human writing.
5Sentence Structure
AI models tend to construct longer sentences and overuse the conjunction 'and'. This is one reason for their reduced punctuation usage. AI texts also contain far fewer expert quotes, making them appear less authoritative. Paragraph structures are more mechanical and predictable compared to human writing.
6Style Variations
The research showed that AI writing style can change significantly with software updates. The evolution in em-dash usage proves this. It demonstrates that AI detection methods need constant updating. While AI style approaches human writing, it still lacks clarity and elegance, often following formulaic patterns.
7Literature Test
Excerpts from bestselling novels published between 1950-2022 showed AI texts have different structures from literary works. AI struggles with creating literary elements requiring character development and emotional depth. This reveals AI's continued limitations in creative writing.
8Conclusions and Implications
The Economist's comprehensive research proves AI-generated content is detectable. However, detection methods require constant updates. For original content creators and editors, these findings serve as an important guide for quality control of AI-generated content. Human writers can produce more authentic works by avoiding these AI characteristics.
Original article: Daring Fireball — Yapay Zeka Yazıları Nasıl Tespit Edilir?