Chatbots have learned to write from us. Can AI change the way we think?

When AI saturates the internet, researchers say it's changing the way we write.

A recent report from online security platform Imperva found that automated traffic was outpacing human-generating activities online for the first time in 2024.

And it suggests that as people turn more and more into AI-powered chatbots in their daily lives, experts tend to use language chatbots. Some people worry about this reducing human creativity and creating feedback loops that could potentially change the thinking process.

“We are concerned that homogenization of coal mining language is a coal mining canary because of the homogenization of thought, not just what we say, but also what we think.

She says that more than half of the online texts are likely to be generated entirely or in part by AI now.

Bovell says he has noticed the uniformity of the way people write on social media platforms such as X and LinkedIn, as well as sites such as Subscack, blogs and newsletter platforms.

Several experts told CBC News that the increasing online uniformity of this language has made it increasingly difficult to analyze things that have not been written by humans.

According to Bovell, AI writing features include symmetric clauses such as the word “not just X but Y,” the word “” like “Y,” the use of “list” and bullet points, often meaningless metaphors, and generally bland neutral tones.

AI also affects human thoughts and reasoning

Morteza Dehani, director of the Center for Computational Language Sciences at the University of Southern California, says his research will reinforce concerns about homogenization of language and thinking.

“We are losing the dispersion and perspective we see in human society, and this will also affect our reasoning,” Dehani said.

In a February survey, Dehghani and a team of USC researchers analyzed the language found in Reddit submissions, scientific papers and American community newspapers from 2018 to 2025.

They found a text spike generated in AI in the second half of 2022. This corresponds to Openai's ChatGPT chatbot release. They also discovered a decline in the variance and complexity of written texts since that spike.

Dehghani says this seems to be trying to adapt to the writing structures seen in an online world where AI is increasingly overrunning, even writers who don't use large-scale language models (LLMs) directly like ChatGpt.

“I want to write in the same way my readers are exposed or used to,” he said. “We're just in this homogenization loop.”

In another paper published this month, Dehghani and other USC researchers argued that these homogenizing effects LLM have on writing are brought into human representation and thinking, saying LLM “reflects and reinforces” the dominant style of writing that “alliminates” “alternative voices and reasoning strategies.”

Futurist Bovel says that the data used to train AI comes from the internet, and therefore tends to reflect the loudest and dominant online voices. This implies historically marginalized groups and cultures, and is not normally used in training data added to homogenization.

This issue is exacerbated by the fact that most of the AI we use comes from a small number of American companies.

“At the end of the day, these companies are 'building the foundations of the future,' she said. That's something we all really need to think about.”

AI is increasingly training on its own content

John Licato, an associate professor of computer science and engineering and director of the Advancing Machine and Human Reasoning Lab at the University of South Florida, says the amount of AI and bot-generated content we are consuming is higher than ever, and will likely continue to grow.

He says it is particularly difficult to determine the exact level of automated and human content because humans amplify posts created by social media bots and vice versa.

He says that the internet has reached the point where it can continue to function in an imaginary scenario where humans have stopped using it.

That leads to another problem with machine learning. There, generation AI systems trained with unique content become increasingly worse and worse, producing more biased results. This means that the likelihood is already happening.

“If you have AI that is continuously trained with your own data, things like bias get worse after multiple generations. Mistakes get worse after all generations,” he said. “That's what you would expect to happen if a person just fell off the internet.”

Promoting to maintain diversity

Those studying AI and language suggest that there is still much to learn about this uniformity and how to deal with it.

In a USC paper published in August, researchers concluded that “preserving and enhancing meaningful human diversity” must be at the heart of AI design and development. If we don't pay careful attention to its diversity, they warn, we cannot harness the full potential of technology “without sacrificing the very diversity that defines human society.”

To oppose the homogenization of LLMS and to work in the public interest, Bovell says it must find ways to form data in a more diverse way.

It's also important to have open source models that anyone can change them, Bovell says it doesn't publish code or other details, as opposed to its own models like CHATGPT. She encourages countries like Sweden have begun working on the “sovereign AI” model to reflect local culture.

Bovell says that if AI influences language, a shared fabric, to enable society to reach consensus and move forward, “I want to make sure these tools reflect the widespread and depth of diversity in your population and there.”