AI doesn't produce speech like caregivers do
Abstract
AI-generated child-directed speech is increasingly positioned as a supplement or replacement for caregiver speech, either for use in early-learning technologies (AI-powered toys) or as a substitute to human data for research. We compare child-directed speech from AI and human caregivers along dimensions linked to child language outcomes: utterance length, lexical diversity/complexity, pronoun use, wh-words, and verb types. While the AI-generated speech is longer than caregiver input, it is less variable and exposes children to fewer word types. AI and caregiver speech also show different distributions of grammatical elements: AI favors first- and second-person pronouns, while caregivers use a broader set of referential forms; AI produces primarily argument wh-words, whereas caregivers use wh-words that appear in more complex syntactic constructions; and AI tends to use different kinds of verbs. These findings suggest that AI may not sufficiently imitate child-directed speech with the properties required for early language learning.