stochastic parrot

"On the Dangers of Stochastic Parrots", by Bender, Gebru, McMillan-Major and Shmitchell, was published at the FAccT conference in March 2021. The label is the paper's own, and the claim it carries is narrow: statistics without meaning, the way a parrot repeats sounds.

The claim is about how the text gets produced, not how good it turns out. No benchmark score comes into it, and nothing in the paper says the output is useless. A reader who hears "stochastic parrot" as shorthand for "the model is stupid" has picked up the wrong target, because a system can stitch statistics without meaning and still produce text you can use.

The term now travels in two opposite registers. The paper used it as a critique. The p(doom) song uses it ironically, which osmarks, who wrote the annotation behind The p(doom) song, annotated, says outright, crediting the ironic reading to a suggestion from a Claude model. Meet the ironic use first and the paper itself reads as a taunt. Meet the paper first and you miss the joke.

The chronology that gets told backwards

The term usually arrives attached to a story about Timnit Gebru's job at Google, and the story is usually told in the wrong order. The annotated essay corrects its own earlier draft on exactly this point: Gebru left Google in December 2020, which is before the paper was published, so the paper cannot have cost her the job in that order.

What the record holds is this. The dispute was over the draft. Google asked for a retraction, or for its authors' names to be taken off it. Gebru refused, and she left in December 2020. She says she was fired; Google said it was accepting her resignation. Both accounts belong in any retelling, because the two sides describe the same departure in incompatible terms. Margaret Mitchell, who signed the paper as Shmitchell, was dismissed about two months later.