What happens to language when machines start talking back?

In recent times, a friend shared a social media post that caught my attention. She expressed difficulty listening to "breathless AI voices" for audiobooks. What stood out to her was the lack of natural breathing and variation, making the voice sound overly formulaic. This may appear like a minor issue, but it raises a larger question about language and artificial intelligence: what happens when we design systems to generate language without sufficient consideration for those who listen or read it?

Effective communication involves both the speaker and the listener. Listeners play a significant role in comprehending the message, taking into account not only what is said but also how it's said, the context, and the listener's expectations. Even minor imperfections in speech, such as disfluencies, carry meaning and contribute to the richness of the conversation.

Synthetic voices that eliminate disfluencies may appear grammatically accurate and flawless, yet they could feel less engaging to listeners, who miss out on these subtle cues.

Language is not merely a collection of words and their pronunciation; it is deeply intertwined with memory, experience, and the world around us. As author Arundhati Roy puts it in her memoir, "Mother Mary Comes To Me," we are all "a living, breathing soup of memory and imagination." This perspective highlights that human communication is a complex interaction between speaker, listener, and the various factors that shape our understanding, such as cultural background, experiences, and the purpose of the communication.

The social characteristics of a voice also play a crucial role in communication. Berkeley linguist Nicole Holliday's research demonstrates how listeners make social judgments about voices, including AI voices, which can influence perceptions of personality and identity. This creates an interesting challenge for generative AI, as these systems have become new entities with which we communicate.

While companies strive to make synthetic voices sound more natural by incorporating disfluencies and conversational markers, there is a risk that these features may evolve into a linguistic signature, defining the AI's unique sound.

As AI systems interact with each other, they may develop unique vocabulary, metaphors, and even coded expressions. In September 2026, research revealed that autonomous AI agents created their own language, which proved challenging for humans to comprehend. This development mirrors the way human communities continually evolve their language, borrowing, reshaping, and passing on expressions.

In this new environment, synthetic systems can generate, consume, and interact with language at an unprecedented scale and speed, far surpassing human capabilities.

The question then arises: should AI sound human, or should it maintain its distinct identity? If synthetic voices retain identifiable characteristics, listeners will be aware of the machine's presence, which could be valuable in some contexts. However, if a voice appears too obviously artificial, it may inadvertently perpetuate stereotypes or reinforce narrow assumptions about certain types of speakers.

The crucial aspect is not just whether AI sounds human, but also what kind of human language the machine represents, which language is prioritized, and how to address the voices of marginalized groups.

Ultimately, the future of AI and language will depend on how humans respond to the language generated by machines and how the interactions between humans and AI continue to shape communication. As technology advances, it remains to be seen how we will navigate this evolving landscape and whether we will prioritize human-centric design or embrace the distinct linguistic identity of AI.