How the Language of AI Could Change Human Speech

On his blog, the noted security expert Bruce Schneier raises the question of how the spread of large language models may affect human speech.

A Narrow Slice of Language

Large language models are trained on written texts — from textbooks to social media posts — as well as on speech captured in films and on television. Yet they have almost no access to the unrehearsed conversations that people carry on face to face or over the phone.

And it is precisely this spontaneous spoken language that makes up the overwhelming bulk of all human communication and forms a vital element of culture. It turns out, then, that the models are trained on only a narrow, atypical slice of language.

Where the Risk Lies

Schneier points to the danger posed by this quirk of training. As the use of language models grows, people increasingly encounter texts generated by AI. Over time, this could lead people themselves to start adopting the linguistic patterns and speech habits characteristic of such models.

In the author's view, this could affect not only how people communicate with one another, but other aspects of language and thought in a broader sense.

Why It Matters

Language is not merely a tool for conveying information; it is the foundation of culture and of a way of thinking. If the mass spread of AI-generated text begins to gradually shift the norms of spoken language toward the patterns typical of language models, it could impoverish the natural diversity of speech — that very spontaneous, unrehearsed communication that the models are as yet unable to reproduce and on which, in essence, they were never trained.

Schneier offers no ready-made solutions, but he draws attention to the fact that the consequences of mass interaction with AI reach far beyond convenience and efficiency — they touch the very nature of human communication.

Source: Schneier on Security