What is AI doing to human language?

By Domini Stuart

Language has never been static. 

Shakespeare introduced about 400 words to the English vocabulary. 

Hip-hop turned slang into common speech. 

We are constantly coining words to keep pace with innovation — think vibe coding and biohacking. But is artificial intelligence (AI) doing the opposite: narrowing human speech into something safe, limited and bland?

“So far, the most visible changes are to vocabulary,” says Dr Ward Peeters, linguistics lecturer and director of the Master of Applied Linguistics at Monash University. 

“For example, words like ‘delve’, ‘realm’, ‘significant’ and ‘underscore’ have spiked sharply [likely due to AI use] in academic and biomedical literature since late 2022 and are more common in popular media and social media posts. Research has also shown that when AI revises texts written by humans, the range of words is reduced.”

Changes in the richness and complexity of language are more subtle.

“There are indications that variations in style have started to shrink across texts and authors, making text in general more homogenous,” Peeters says. 

“Tone of voice, the way you structure your sentences and your use of hedging words like ‘perhaps’ or ‘seems’ are not just decorative. They are ways of building trust, establishing expertise and positioning yourself in relation to your audience. 

“When AI revises that towards a neutral, pleasant, generalised norm, it weakens your communicative scaffolding.” 

AI may also impact cognition and creativity. 

“Language helps us to shape what we think, so when the language we are using becomes limited and predictable, our ability to reason and innovate is likely to follow the same path,” says Simon Lesch, APAC AI and transformation lead at Archetype, a strategic PR and innovation consultancy. 

“And, while we tend to think of creativity in terms of things like music and art, business is a hugely creative field. It demands constant invention — new solutions, new narratives and new ways of making sense of change.” 

A study by the MIT Media Lab asked subjects to write an essay in one of three ways: using ChatGPT, a search engine or with no help at all. 

Those using ChatGPT showed significantly less engagement in the frontal regions of the brain, which are the areas most associated with critical thinking, planning and memory. When they were later asked to write without AI assistance, their neural connectivity remained weaker than those of the control groups. 

 

"There are indications that variations in style have started to shrink across texts and authors, making text in general more homogenous. Tone of voice, the way you structure your sentences and your use of hedging words like ‘perhaps’ or ‘seems’ are not just decorative. They are ways of building trust and establishing expertise."

 

“It is early work and the sample was very small — there were 18 participants in each group and only nine returned for the crucial final session,” Peeters says. 

“However, the direction of the findings is consistent with cognitive load theory, which has long predicted that when you repeatedly remove the kind of effort that produces deep learning, you stop building the capacity. 

“The researchers coined the term ‘cognitive debt’, but I prefer to use ‘cognitive offloading’, as it is more widely understood,” he says.

Lesch sees cognitive offloading exemplified in GPS navigation.

“Back in the day, you used a map,” he says. “Now, the GPS does all the work for you and, while I have no doubt that we are arriving at our destinations faster and more easily, we have largely lost the ability to get there by ourselves.”

Cognitive offloading could also take a serious toll on business.

“If junior analysts and accountants are drafting documents and interpreting data with a great deal of assistance from AI, they may not be building the professional judgement that comes from doing that work themselves,” Peeters says. “The pipeline of expertise could thin out that way.” 

The great ‘linguistic flattening’

With AI, prediction and averaging can combine to flatten language and cognition.

“Large language models (LLMs) are trained on vast quantities of text to predict the next word or a small chunk of text known as a ‘token’,” Lesch says. 

“When AI completes ‘The cat sat on the…’ with ‘mat’, it has no concept of cats, mats or sitting. It has simply identified the most likely sequence based on billions of examples.”

Similarly, if you give AI a vague prompt such as “write a press release”, the model will fall back on the patterns it sees most frequently. 

“This nudges the language towards the average — something that feels accurate, acceptable and safe,” Lesch says. 

 

"Large language models (LLMs) are trained on vast quantities of text to predict the next word or a small chunk of text known as a ‘token’. When AI completes ‘The cat sat on the…’ with ‘mat’, it has no concept of cats, mats or sitting. It has simply identified the most likely sequence based on billions of examples."

 

“Averaging has some benefits, for example, it could help non-native speakers to increase their vocabulary. The problem is that if you are a more sophisticated writer, it is likely to reduce your vocabulary and blunt your style.”

AI also comes with intrinsic biases.

“LLMs are trained on huge data sets mainly scraped off the internet,” Peeters says. “That sounds comprehensive until you consider what it means. The web is not a neutral sample of human language. It over-represents English, Western (specifically American) cultural norms and favours the written over the spoken, the formal over the vernacular and the prolific over the considered.

“This generates plausible sounding text with remarkable fluency, which we find psychologically reassuring even when the content is wrong,” he continues. 

“For example, a financial professional who relies on AI-drafted language for a regulatory filing may not notice that fluent and convincing language is hiding a term that is subtly different from its technical definition.”

How to use AI and keep the sound of authenticity

There are ways to prevent AI from flattening thought and language.

“You can specify things like your preferred tone and structure, your target audience and things you particularly want to avoid,” Lesch says. “When I am giving instructions to AI, I include a list of words I do not like or that are associated with AI. 

“I tell it to consider my context, such as how old I am and where I live. I also write the first draft myself, even if it is just bullet points or rough paragraphs. If you start out with a draft generated by AI, it will always be generic.”

 

_{Production credits
Banner image by LumiNola via Getty Images}