AI English Falls Short on Variation and Readability
Get the Health newsletter
Daily health & science — research, biotech, public health, the studies worth knowing. Free.
- A linguistics researcher identifies variation and readability as the two qualities distinguishing human English from AI English, which the author calls 'exam English' — formal, dense, and perceived as both robotic and smart.
- A side-by-side text message test demonstrates the contrast: the human version uses lowercase, phonetic spellings ('ur,' 'u'), and varied clause structures, while the ChatGPT-generated response uses formal capitalization, commas, and a repeated clause pattern.
- GPT 5 was 'infamously cold-sounding' compared with its predecessor GPT 4, and the patterns favoring homogeneity and information density are likely to persist, the researcher argues.
- Instructional tuning — training AI models to follow human instructions — only makes AI English less like human English, not more, the researcher contends.
- The article recommends describing passages with language labels like 'dense' or 'plain' rather than social judgments like 'sounds smart,' and using AI tools selectively to prevent AI English from crowding out human variation.
- Tools like small language models, Bias Shield, and translingual chatbots can help users make principled choices and bring more of the global variation in human English to AI outputs, the researcher notes.
Why it matters: Users who default to AI English risk producing writing that is less personable and readable, especially for diverse audiences, since the models are trained on a narrower band of English than humans actually speak and write — making AI literacy about pattern recognition, not just fact-checking.
