ChatGPT's language comprehension exposed: a look at its linguistic depths
In a study published in PLOS One, psycholinguist Michael Vitevitch from the University of Kansas explored how AI, specifically ChatGPT, processes language, particularly when presented with complete linguistic nonsense[1][2].
The research involved feeding ChatGPT nonwords - meaningless letter and sound combinations often used in cognitive psychology to probe language processing. When presented with these nonwords, ChatGPT excelled at pattern recognition, but its approach was found to be different from human linguistic cognition.
Unlike humans who rely on phonological, semantic, and contextual cues informed by their language experience, ChatGPT uses statistical associations learned from its training data to find patterns[1][2]. This means that while ChatGPT can identify patterns in nonsensical language inputs, it does not process them using the same mental mechanisms humans do.
When humans encounter nonsense, they often try to impose meaning or follow phonetic rules based on their internalized language knowledge. However, ChatGPT's approach is more about probabilistic pattern matching without true comprehension or semantic grounding[1].
In contrast to human responses, ChatGPT's inventions were often predictable, relying on a method of combining two existing words to create new ones[1]. One of the more interesting creations was 'rousrage,' for anger expressed upon being woken.
The study also revealed that ChatGPT is not always accurate when defining extinct words, trying to be helpful in some instances by hallucinating definitions[1]. For instance, when asked about 'upknocking,' a 19th-century job where people tapped on windows to wake others before alarm clocks, ChatGPT made up a definition.
Despite these differences, Vitevitch argues that the goal is not to mimic human cognition, but rather to identify where AI can complement our linguistic strengths[1]. He believes that understanding these differences between AI and human language processing is crucial for the future development of AI and its potential applications.
The research, originally published by Cosmos under the title "What nonsense reveals about ChatGPT's understanding of language," provides valuable insights into the unique ways in which AI processes language, emphasizing both the strengths and limitations of AI language comprehension[1][2]. This difference highlights that human and AI approaches to language are distinct but can be complementary.
[1] Vitevitch, M. (2023). What nonsense reveals about ChatGPT's understanding of language. PLOS One. [2] Cosmos. (2023). What nonsense reveals about ChatGPT's understanding of language. Retrieved from https://cosmosmagazine.com/technology/what-nonsense-reveals-about-chats-understanding-of-language/
The study discovered that unlike humans, ChatGPT, a form of artificial intelligence, does not process nonsensical language using phonological, semantic, and contextual cues but instead relies on statistical associations learned from its training data for pattern recognition.
Furthermore, ChatGPT's inventions, such as 'rousrage' for anger expressed upon being woken, are often predictable and result from a method of combining two existing words to create new ones, indicating its unique approach to language generation, which differs significantly from human language processing.
