Five years after the publication of the influential paper “On the Dangers of Stochastic Parrots,” computational linguist Emily Bender says the phrase has taken on a life of its own, often in ways that misrepresent its original intent. In an interview with IEEE Spectrum, Bender explains that the metaphor was never meant to dismiss every advance in large language models (LLMs). Instead, it was created to highlight a fundamental limitation: LLMs generate text by identifying statistical patterns in language rather than by understanding the meaning behind the words they produce.
Bender also challenges the widespread use of the term artificial intelligence. She argues that it groups together very different technologies under a single label while exaggerating their capabilities. According to her, clearer descriptions of specific technologies, such as language models, machine translation, or speech recognition, lead to more informed discussions about regulation, deployment, and public expectations. She emphasizes that computational linguistics is a field dedicated to understanding language and building useful language technologies, not necessarily pursuing artificial general intelligence.
A major concern raised in the interview is the tendency to attribute human qualities to chatbots. Because LLMs produce fluent, conversational responses, users often assume they possess reasoning, intentions, or comprehension. Bender argues that this perception is misleading. The models generate plausible text by learning statistical relationships from vast datasets, not by connecting language to lived experience or an internal understanding of the world. This distinction becomes especially important when these systems are used in high-stakes applications such as education, healthcare, legal services, or public information.
The interview also revisits the broader themes of the original paper, including the environmental cost of training massive models, the biases embedded in training data, the opacity of commercial AI systems, and the risks of deploying language models without sufficient transparency or accountability. Bender argues that these issues remain just as relevant today, despite rapid improvements in AI performance.
Rather than rejecting language technology, Bender advocates for a more precise understanding of what these systems can and cannot do. The article serves as a reminder that meaningful discussions about AI require careful language, realistic expectations, and continued attention to the ethical and technical challenges surrounding increasingly capable language models.