Stochastic parrot

“Stochastic” means governed by probability; a parrot repeats what it has heard. The phrase was introduced by Emily Bender, Timnit Gebru, Angelina McMillan-Major and Margaret Mitchell (writing as Shmargaret Shmitchell) in a 2021 paper on the risks of ever-larger language models. Their claim was that a language model stitches together linguistic forms it has observed, guided by probabilistic information about how they combine, and that the coherence readers find in the output is supplied by the readers.

What the label claims

Two claims travel under the name, and they differ in strength. The first is about training: a model learns from the form of text alone, with no access to what the text is about. The second is about the result: so it cannot mean or understand anything. The first is a description of how such systems are built. The second is an inference, and it is the contested one. It belongs to the same family as Searle’s Chinese room, which argues that manipulating symbols by rule is never sufficient for understanding.

The replies

Critics answer that learning to predict text well may require building internal models of what the text describes, and point to research finding structured representations of space, time and truth-value inside trained networks. Others accept that models lack human-like understanding and deny that this settles what they can do.

Where I stand

I do not rest anything on the claim that language models understand nothing, and I think the label, used as an insult, proves less than its users hope. The question I find more tractable is narrower: when a model asserts something, does the truth of that particular claim figure in why it was produced? That is the Tracking Minimum, and on current evidence these systems do not meet it for open-domain assertions. A system could fall short of that condition without being a parrot, and a parrot, strictly, never asserted anything at all.