The stochastic parrot is the invention of linguist Emily Bender – the creator of another memorable fictional animal, the Statistical Octopus  – which she uses to criticize the Large Language Models (LLMs) on which ChatGPT and other forms of AI are based. As a reminder, “stochastic” means “determined by random, probabilistic distribution.” A stochastic parrot, says Bender, is an entity “for haphazardly stitching together sequences of linguistic forms (…) according to probabilistic information about how they combine, but without any reference to meaning.” This, she says, is exactly what LLMs do.

The stochastic parrot first appeared in a March 2021 article which Bender published with three co-authors, two of whom, Timnit Gebru and Margaret Mitchell, were co-leads at Google’s Ethical AI team. In this paper, they discussed:

• The biases inherent to LLMs due to the gender, racial, and political biases implicit in the texts used to train them.
• The practical impossibility of examining the data used to train the models, as it comprises billions of words.
• The environmental damage caused in order to physically power LLM processing.
• The rapid obsolescence intrinsic to using language from a specific period in time.

Interestingly, Sam Altman, the CEO of OpenAI, the company that created ChatGPT, embraced the term, famously tweeting “i am a stochastic parrot, and so are u”.[1] In arguing  that human beings are indeed a form of stochastic parrot, he posited that, in the last term, our linguistic utterances – and possibly all our choices – are probabilistically determined. Thus there would be little to differentiate our language, and arguably our overall behavior, from that generated by an AI, thus making AI potentially indistinguishable from human beings (as required by the famous Turing  Test).

This, Bender argues, is self-evidently ridiculous: people are not parrots and human speech is not probabilistic. LLMs, Bender says, are “machines that can mindlessly generate text”, whereas human beings do have minds – we have cognition, understanding, whereas an LLM no more understands the texts it is processing than a toaster understands the toast it is making.

Furthermore, according to Bender, equating the way in which human beings and LLMs generate language is tantamount to anthropomorphizing technology, in so doing reducing  the complexity of human language. “People want to believe so badly that these language models are actually intelligent,” she points out, “that they’re willing to take themselves as a point of reference and devalue that to match what the language model can do.”

The scientific journal Nature  has stated that “conversational AI is a game-changer for science” as a super-powered search engine. But when it comes to original writing, there is a caveat: “there is a huge gap between writing funny instructions for removing food from home electronics and doing scientific research.” 

Matthew Salganik , a Professor of Sociology at Princeton University’s Center for Information Technology, used ChatGPT to assist him in writing peer reviews. He used two types of prompts to improve a review he had already written: aesthetic and technical. The results were discouraging.

“For both kinds of prompts, ChatGPT didn’t produce anything useful to me. Instead, it often avoided aesthetic questions by providing summaries of the paper (which wasn’t very useful because I had already read the paper carefully). For technical questions, ChatGPT avoided my question by defaulting to more general questions about common problems with statistical analysis. These general answers were not bad, they were just not helpful for improving my review. After about 20 minutes of trying and failing to get something interesting, I gave up and submitted my review with no changes.  After almost 90 minutes of extra work, my review was not improved one bit.”

It seems that, whatever claims pet-shop keepers may make, the stochastic parrot’s inadequacies are of a rather permanent nature – it’s not just that it’s pining for the fjords.

–

[1] https://twitter.com/sama/status/1599471830255177728?lang=en