AI-generated text exhibits a statistically significant preference for voiceless and alveolar consonants and a reduced frequency of dental and labiodental consonants compared to human-written text, indicating that orthographic-to-phonological modeling in large language models introduces systematic biases in consonant distribution that differ from human speech patterns.

When a system learns to convert written words into spoken sounds from text data, it repeats the most common letter-sound patterns it sees, which are often from books and websites. These patterns favor sounds made with the tongue tip and without vocal cord vibration, like 't' and 's', because they appear more often in written text. Sounds made with the lips or between the teeth, like 'v' and 'th', are less common in writing, so the system uses them less. This creates a speech-like pattern that differs from how real people naturally speak.