New research indicates that popular AI chatbots, including OpenAI's ChatGPT and Google's Gemini, have developed unique and identifiable writing styles, or "idiolects," akin to human speech patterns. This discovery, stemming from linguistic analysis, suggests that these large language models don't merely average their training data but form their own consistent linguistic habits.
Linguist Karolina Rudnicka employed computational methods to analyze hundreds of texts about diabetes generated by both ChatGPT and Gemini. Using the Delta method, a standard technique for authorship attribution, she found clear stylistic differences between the models. For instance, ChatGPT's texts showed a linguistic distance of 0.92 to other ChatGPT content, while their distance to Gemini content was 1.49. Similarly, Gemini texts had a distance of 0.84 to other Gemini content and 1.45 to ChatGPT, confirming distinct authorship patterns.
The analysis revealed specific stylistic preferences:
ChatGPT tends to favor more formal, clinical language, frequently using phrases such as "individuals with diabetes," "blood glucose levels," and "characterized by elevated." It also uses "glucose" more than twice as often as "sugar." ChatGPT was observed to overuse sophisticated verbs and adjectives like "delve," "align," and "underscore."
Gemini, in contrast, adopts a more conversational and accessible tone, employing phrases like "high blood sugar" and "blood sugar control." "High blood sugar" appeared 158 times in Gemini's dataset compared to just 25 times in ChatGPT, while "blood glucose levels" appeared only once in Gemini's entire dataset.
Further examination of trigrams (three-word combinations) also highlighted consistent sentence structuring patterns for each model. This research has significant implications, particularly for educators and teachers, who could potentially use these identified patterns to detect AI-generated submissions from students.
This groundbreaking research underscores that AI chatbots are not just replicating their training data but are developing distinctive linguistic fingerprints, mirroring the complexity of human language development.
Sources:
No comments:
Post a Comment