Bender (co-author of the 2021 Stochastic Parrots paper) corrects five years of accumulated misreadings: the phrase is a description of LLMs, not a critique, hypothesis, or insult.
Key Takeaways
“Stochastic parrot” describes LLMs as systems that remix linguistic form without grounded meaning or communicative intent – not a ranked judgment about capability.
The paper’s actual target was human behavior: data theft, undocumented datasets, exploitative labor, and environmental cost – not the models themselves.
The “it’s come up with something new” objection misses the stochastic part: outputs are probability-shaped remixes of training data, not regurgitations.
Multimodal (image/text) models may meet a thin technical definition of understanding from Bender & Koller 2020, but the stochastic parrot framing still applies to how users interpret their language output.
Deployed systems like ChatGPT, Claude, and Gemini may route queries to calculators or classifiers – the LLM is one component, not the whole product.
Hacker News Comment Review
Sparse discussion; the main tension is between the article’s framing of LLMs as non-thinking mimics and the empirical reality that current models achieve things like IMO gold – commenters flag that “merest resemblance of thinking” undersells observed performance.
Notable Comments
@CamperBob2: pushes back on dismissive framing – “the merest resemblance of thinking is enough to take gold at IMO.”
