A new study has warned that large language models (LLMs) like ChatGPT, widely used for writing and decision-making, are reducing human cognitive diversity. By favouring common patterns during interactions, these LLMs are making language and ideas more uniform, suppressing unique voices and perspectives, which could gradually diminish creativity.
“With the growing prevalence of LLMs across fields and in shaping communication, studies have found that they are influencing and homogenising people's language, perspectives, and reasoning,” the report titled - The homogenising effect of large language models on human expression and thought - noted.
Also Read | AI Backfire: Artificial Intelligence Usage Can Lead To Burnout, Low Quality Work, Study Finds
Published in the journal ‘Trends In Cognitive Sciences' on Wednesday, the report warned that as LLMs become more widespread, they could subtly influence how people think and communicate.
“This homogenisation arises not only from biases in the training corpora but is also amplified through the training process itself, favouring patterns that are frequent and easily generalisable while smoothing over minority representations,” the paper argued.
Use Of AI May Reduce Our Ability To Adapt
Authors of the paper worry that losing diverse ways of thinking and communicating could limit problem-solving and reduce our ability to adapt to new situations. Zhivar Sourati, a computer scientist of the University of Southern California and first author for the paper, told CNET: “Diversity of language, perspective, and reasoning isn't just a cultural nicety; it's functionally essential. It's what drives creativity, innovation and collective problem-solving.”
According to Sourati, this trend affects even those who don't use chatbots. This is because people may feel pressured to align their thinking and speaking with others, seeing it as more socially acceptable. “The writing styles that LLMs incorporate are not neutral but a source of bias themselves, reproducing dominant expressive norms while eroding minority and underrepresented voices,” the paper warned.
Also Read | Economics Of AI Depends On Nuclear Power | The Reason Why
LLM Development Methods Need Re-evaluation
It also noted that a number of researchers have suggested ways to reduce LLMs' tendency to make language and ideas uniform.
“Prompting-based methods, applied at inference time, aim to enhance output diversity by modifying how prompts are phrased or conditioned. Training-based approaches instead modify the LLMs' learning process to optimise for semantic and stylistic diversity. Although promising, evidence of homogenisation in sociolinguistic contexts calls for closer evaluation of whether these strategies foster genuine diversity,” it said.
To address this brewing crisis, the paper emphasised that preserving and enhancing meaningful human diversity must guide LLM development and evaluation.
“Only through deliberate attention to this pluralism can we harness the full potential of language technologies without sacrificing the very diversity that defines human society,” it noted in concluding remarks.
Essential Business Intelligence, Continuous LIVE TV, Sharp Market Insights, Practical Personal Finance Advice and Latest Stories — On NDTV Profit.