When too many texts published online, in newspapers or on social media are written, sometimes even thought up, by AI, something happens to the language we all share. The innocent and largely unacknowledged victim is the richness of the language through which we understand, describe and discuss reality.
Historically, the press has been one of the spaces in which public language grows, becomes richer and spreads new words, expressions and ways of describing events. In many nations, newspapers have shaped and standardised the population’s everyday language. But if a significant share of writing is delegated to generative AI, this function is weakened.
Large language models produce fluent, credible texts. By now, everyone knows that, drawing on a vast statistical base, they predict the most probable next word or token. And even if they do so convincingly, this leads them to favour statistical regularities, common formulations, well-established arguments and predictable structures. This is why a reasonably attentive reader immediately notices whether a post or an article has been written by AI. They may then carry on reading for the information it offers, but with that awareness and irritation. This is because AI writes in a predictable, recognisable and even slightly disturbing way. Disturbing because when an empty-shell form becomes dominant, the sense of the content is lost and all one reads is the form. And the immediate effect is a diminished capacity to convey meaning and information.
What is more, the risk increases when models begin training on texts already produced by other models. This is model collapse: a degenerative process in which synthetic material contaminates subsequent training data. If public spaces fill up with artificial texts, the future linguistic ecosystem becomes narrower, less exposed to the social, stylistic and pragmatic variety of human language.
Statistics and regression to the mean mean that, for example, linguistic structures dwindle, biases grow stronger, writing loses the specificity of unusual words, and registers become standardised, losing irony, ambiguity and nuances of perspective.
Regression to the mean is what happens when the floor is raised but the ceiling is lowered: the system is better at preserving the centre of the language, its ordinary part, and loses literary usages, brilliant departures from the norm and specific words for specific phenomena. To be clear, Calvino or Gadda would not exist. Nor would Manzoni.
And we could ignore this phenomenon if it concerned only the performance of the models, but the distributed linguistic ecosystem is being caught in the crossfire.
We can already see that educated people who used long dashes – to frame an elegant aside – no longer do so now that they have become AI’s hallmark, at least as much as the rocket-taking-off and little target emojis that still accompany posts on LinkedIn. And we also see, especially in video reels, people reciting texts written by ChatGPT from memory. So we find ourselves having to listen to people saying: "So-and-so’s brutal act was neither a mistake nor an accident. It was a choice that invites us to reflect on the importance of empathy in our social fabric."
I do not mean that every time someone copies and pastes from ChatGPT we move towards utter decline: when synthetic texts are mixed with exclusively human thoughts, rewritten or supplemented, there is practically no harm. But today it seems that human writing is being replaced en masse and then reused as though it were living language.
Writing had two functions: the direct meaning in what it said and the indirect meaning in the language it gathered, created, reworked and disseminated. Now it seems to have relinquished this second function, losing one of its most important cultural roles.