Spanish scientists from Rovira e Virgilia University and the University of Barcelona have discovered that the most advanced artificial intelligence (AI) systems based on large model languages are significantly inferior to humans at recognizing grammatical errors. The study was published in the scientific journal magazine Proceedings of the National Academy of Sciences (PNAS).
Experts tested three of today’s top language models, including ChatGPT. They were given a simple task: to determine which sentences were grammatically correct and which contained grammatical errors. A group of people completed the same task.
While live participants easily detected grammatical inconsistencies, the AI provided many incorrect answers. In fact, artificial intelligence correctly detected all of the given examples.
According to the authors of the study, the experiment showed that artificial intelligence cannot yet evaluate texts for compliance with grammatical norms, even if these rules are loaded. However, one notices insignificant errors in grammar instantly and effortlessly. This suggests that neural networks understand language differently than humans.
Previous scientists I learnedthat some AIs are biased against men and women.