AI is making us sound the same—and killing our personal expression
More than a third of all internet web pages published since ChatGPT’s November 2022 launch were authored by AI , according to the Pew Research Center . And the writing is becoming simpler than ever. That flood of AI-generated language may also be changing the way humans write. A new paper in Nature Human Behavior finds that LLMs can push people’s writing toward a narrower set of linguistic norms,…
Since the advent of ChatGPT in November 2022, over a third of all internet content has been produced by artificial intelligence, according to the Pew Research Center. This surge in AI-generated writing is simplifying language use and potentially altering how humans express themselves. A new study published in Nature Human Behavior reveals that large language models (LLMs) can standardize writing styles, diminishing linguistic diversity and diminishing cues that signal individual identity.
Lead researcher Zhivar Sourati recalled reading James Pennebaker's book on the influence of personal values and backgrounds on word choice when LLMs gained popularity three years ago. He questioned whether these models could reduce the diversity found in human language. To investigate, Sourati analyzed approximately 80,000 academic papers, 400,000 news articles, and 300,000 Reddit posts before and after ChatGPT's launch.
He discovered that post-AI adoption, writing complexity became more uniform across all three datasets. A subsequent experiment involved using GPT-3.5, Gemini, and Meta's Llama 3 to rewrite human-authored texts. While the models maintained the original meaning, they reduced writing complexity by 21% to 50%. Sourati then focused on the societal implications by examining how LLMs might obscure individual identity signals in writing.
By using texts from people whose traits were determined through psychological questionnaires, his team found that predictive accuracy for personal traits, such as demographics, personality, empathy, and moral values, decreased by an average of 6% after AI processing. The erasure of linguistic signals varied depending on the trait analyzed.
Sourati warns that relying on LLMs for all writing tasks could have dangerous consequences, as problem-solving abilities, crucial in today's world, depend on cognitive diversity as well as linguistic diversity. He believes that the homogenizing effects of LLMs could be particularly detrimental to those still developing cognitive skills, especially students.
Sourati emphasizes the importance of fostering independent thinking before utilizing LLMs more creatively. He suggests that once individuals learn to reason through problems, they can harness the models' strengths without succumbing to their potential pitfalls on collective problem-solving abilities.
Written by urgent.news from Fast Company's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
