According to a recent report published by the Stanford University research team in the journal "Patterns" under Cell Press, large language models (LLMs) have been widely applied in various formal written communication scenarios. The study analyzed English texts from major platforms such as Newswire, PRWeb, and PRNewswire, and found that a large number of these texts show obvious AI traces.

Robot AI Writing AI Education

The research team used a statistical model called the "Distributional LLM Quantification Framework" to analyze the overall distribution of language features in the corpus, estimating the proportion of texts with AI-generated characteristics during specific periods. It should be noted that this method is currently only applicable to English corpora, so the results reflect only the proportion of English texts with AI characteristics, not that the entire article was "completely written by AI."

The report data shows that the research team analyzed a total of 687,000 consumer complaints, 537,000 corporate press releases, 304 million job postings, and 16,000 United Nations news articles from last year, with the following notable results:

  • Corporate press releases: About 24% of the documents showed AI traces.

  • Consumer complaints: Approximately 18% of the complaint texts had AI characteristics.

  • United Nations news articles: About 14% of the content was identified as "almost directly generated by AI large models or heavily rewritten."

  • Job postings: Nearly 10% of the job content showed AI traces.

Researchers pointed out that LLMs have deeply penetrated formal text writing in various fields of society and have gradually become an important auxiliary tool for people's daily written expression. In specific areas, the study found that the proportion of complaints written by AI in regions with lower education levels reached as high as 19.9%, significantly higher than 17.4% in regions with higher education levels. In job postings, small and medium-sized enterprises or startups are more inclined to use large models to generate content, with approximately 10% to 15% of job announcements containing clear AI traces.

Researchers also emphasized that with the continuous advancement of generative AI technology, its writing style is becoming increasingly similar to human writing. Therefore, it will become increasingly difficult to accurately identify the extent of AI's involvement in text creation in the future.