Before Emily Bender and I have looked at a menu, she has dismissed artificial intelligence chatbots as “plagiarism machines” and “synthetic text extruders”. Soon after the food arrives, the professor of linguistics adds that the vaunted large language models (LLMs) that underpin them are “born shitty”.
Since OpenAI launched its wildly popular ChatGPT chatbot in late 2022, AI companies have sucked in tens of billions of dollars in funding by promising scientific breakthroughs, material abundance and a new chapter in human civilisation. AI is already capable of doing entry-level jobs and will soon “discover new knowledge”, OpenAI chief Sam Altman told a conference this month.
According to Bender, we are being sold a lie: AI will not fulfil those promises, and nor will it kill us all, as others have warned. AI is, despite the hype, pretty bad at most tasks and even the best systems available today lack anything that could be called intelligence, she argues. Recent claims that models are developing a capacity to understand the world beyond the data they are trained on are nonsensical. We are “imagining a mind behind the text”, she says, but “the understanding is all on our end”.