Over the last year, the AI bot ChatGPT has dazzled people with its ability to answer questions, write essays and even code software. Among 13- to 17-year-olds in the United States who have heard about ChatGPT (which is most of them), 19 percent say they’ve used it to do schoolwork. ChatGPT and other chatbots like Bard and Meta AI are all based on large language models, or LLMs for short. These models were trained to craft remarkably humanlike language by being fed vast amounts of text from the internet. And while that text includes Louise Glück poems, Oprah’s Favorite Things gift guides and articles from the New York Times, it also includes, as we know all too well, content that is false, defamatory, violent and horrifying.
As a safeguard, creators of major chatbots have also trained them to refuse to provide inappropriate or harmful information, say, step-by-step instructions on how to steal someone’s identity. But the training is not foolproof, and people have already exploited chatbot weaknesses.
In this issue, physics and senior writer Emily Conover digs into computer scientists’ efforts to keep chatbots on the straight and narrow. It’s a huge challenge, Conover explains, in part because these LLMs are still so new, and scientists are just starting to learn about the chatbots’ vulnerabilities. And the challenge will become much bigger as LLMs are integrated into everyday products or given tasks like running subway systems.
The reality is that even though LLMs sometimes sound human, they aren’t. In reading Conover’s article, I learned the delightful term “stochastic parrot.” Computational linguist Emily Bender of the University of Washington and colleagues use it to explain that while LLMs can compile words into prose, they don’t understand the meaning of what they “write,” and thus can’t understand if it’s inaccurate or immoral. They’re parroting.
Real parrots, and the scientists who study them, may take offense…
Read the full article here