On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?

Recent large language models perform impressively on benchmark tasks. Bender and colleagues argue, however, that scaling alone is unlikely to produce trustworthy language technology, and that responsible development will in most cases require careful documentation of training data, evaluation beyond overall accuracy, and clear attention to who is harmed. The evidence for this conclusion comes from their analysis of systems trained on unfiltered web-scale data, which shows that such systems reproduce — and often amplify — the demographic biases in their training data, produce fluent text that is often factually wrong, and create large environmental and financial costs that fall most heavily on communities that benefit least from the resulting systems. These findings support the conclusion because meaning is grounded in communicative intent rather than in surface form, so a system that only models surface co-occurrence cannot, on its own, be said to understand — and a system that does not understand, but produces fluent output that users mistake for understanding, cannot be trusted at scale. This warranting principle is itself supported by a long tradition in linguistics and the philosophy of language. These conclusions, however, come with limits: the analysis predates the latest generation of instruction-tuned and retrieval-augmented systems, focuses mainly on English-language deployments, and draws largely on case studies rather than controlled experiments. So the specific numbers — though not the underlying mechanisms — may shift as architectures and training pipelines change.