Artificial intelligence is increasingly encountered not merely as a technical tool for information retrieval or task performance, but as an entity that converses with human users. AI-based large language models are now presented as coaches, therapists, and companions of various kinds, and new devices are being developed to accompany particular groups—such as elderly people—in order to alleviate loneliness and compensate for the absence of human relationships. Regardless of how one evaluates these developments, they raise a fundamental philosophical question concerning the status of digital conversational partners. Traditionally, the verb "to speak" is attributed to a who rather than to a what. Does this distinction still hold when we say that "AI speaks"? One may argue that AI does not genuinely speak, but merely combines and rearranges words and phrases originally spoken by human beings and incorporated into its training data. From this perspective, the uniqueness of human speech can be preserved: a who speaks, whereas the device itself - the what - "speaks" only in a metaphorical sense.