‘Chatbots are like parrots, they repeat without understanding’ – Technologist
Emily M. Bender is a linguist, professor and director of the Computational Linguistics Laboratory at the University of Washington. In March 2021, she – along with ethics researchers Timnit Gebru, Angelina McMillan-Major and Margaret Mitchell – wrote an article entitled “Stochastic Parrots”, which has become infamous in the artificial intelligence (AI) sector. The text warned of the limits and risks associated with large language models, software since popularized in conversational bots such as ChatGPT. Today, Bender remains highly critical of the evolution of the AI sector, whose “hype” she rejects in a podcast with sociologist Alex Hanna.
With hindsight, do you think you were right in your 2021 article about the risks of large language models?
I often rather get asked how we feel now that our predictions have come true. And my answer is that those weren’t predictions, they were warnings. And so it feels bad to have observed the beginning of this race for a larger scale of language models, and also what could go wrong, and then to have witnessed people pursuing it anyway.
But there were a couple of things that we missed: we were not tuned in to just how exploitative the labor practices are behind creating these systems [human contractors are paid to annotate data, rate answers or moderate problematic AI content]. Also, we were not aware of just how excited the world was going to be about synthetic text.
You criticize the race towards ever-larger models, but that’s also what has made them better. Should they not have been developed in the first place?
I don’t think it’s clear that they’re better. They are better at mimicking human-sounding text. But I don’t know what that’s for. There are not clear evaluations showing that for one particular task, it gets us better results.
Language modeling is an old technology, that goes back to the work of Claude Shannon in the 1940s. In their original use, language models are an important part of automatic transcription systems or spell checkers or machine translation… But we have way overshot the scale of training data required to do well on those tasks.
Also, if we’re going to build reliable technology, we have to know what’s going into it. And already back in 2020, the scale of training data had passed that point.
According to you, large language models can neither think nor understand, and are “stochastic parrots”. What do you mean by this?
The verb ’to parrot’ means to repeat without understanding. Stochastic parrots is a whimsical metaphor to make vivid a thing very clear to a linguist : languages are systems of signs or symbols. This goes back to Ferdinand de Saussure, the French linguist, from the early 1900s. So there’s always a pairing of form and meaning.And these systems are trained only on the form side. They have no access to meaning and no communicative intent, nothing that they are trying to express. It looks like language, but it isn’t.
You have 73.63% of this article left to read. The rest is for subscribers only.