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

Dehghani says this indicates that even writers who aren’t directly using large language models (LLMs) like ChatGPT seem to be trying to adapt to the writing structures they see in an online world increasingly overrun by AI.

You want to write in the same fashion that your readers are exposed to, or are used to, he said. We’re just getting into this loop of homogenization. 

In a separate paper (new window) published this month, Dehghani and other USC researchers argue that these homogenizing effects of LLMs on writing are carrying into human expression and thought, noting that the LLMs reflect and reinforce dominant styles of writing while also marginalizing alternative voices and reasoning strategies.

Bovell, the futurist, says because data used to train AI comes from the internet, it tends to reflect the loudest and most dominant online voices, meaning groups and cultures that are historically marginalized typically aren’t used in that training data, which adds to the homogenization.

Enlarge image (new window)

This problem is compounded, she said, by the fact that most AI we use comes from a handful of American companies.

At the end of the day, she said, these companies are building the foundation of the future, and that’s something that we all need to really think about.

AI increasingly training on its own content

John Licato, associate professor in 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’re consuming is much higher than it’s ever been, and will probably continue to increase.

He says that determining the exact level of automated content versus human content is especially hard because humans amplify posts made by social media bots, and vice versa.

WATCH | Your next job interview could be with AI:

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

That leads to another problem with machine learning, where generative AI systems trained on their own content produce increasingly worse and more biased results. This is something Licato says is already happening.

When you have AI that’s continually trained on its own data, things like biases get worse after multiple generations. Mistakes get worse after every generation, he said. That is the kind of thing that we would expect to happen if humans just dropped off the internet.

The push to preserve diversity

Those who study AI and language suggest there’s still a lot to learn about this homogeneity and how to address it.

In the USC paper published in August, the researchers concluded that preserving and enhancing meaningful human diversity has to be central to the design and development of AI. If we don’t pay deliberate attention to that diversity, they warn, we won’t be able to harness the full potential of the technology without sacrificing the very diversity that defines human society.

To push back against homogenization in LLMs and make them work in the public interest, Bovell says we have to figure out how to mould data in a way that’s more diverse, something that’s not always a priority for private companies.

It’s also important to have open-source models that allow anyone to modify them, Bovell says, as opposed to proprietary models like ChatGPT that don’t make code and other details available to the public. She’s encouraged that countries like Sweden (new window) have begun working on “sovereign AI (new window)” models to reflect their local cultures.

Bovell says that if AI is going to impact and shape languages that are the shared fabric allowing society to reach consensus and move forward, you want to make sure that these tools are reflective of the people in your population and the breadth and depth of the diversity there.

LISTEN | The decline of the internet (new window)