Why AI Like ChatGPT Struggle with Negation
Artificial Intelligence (AI) has come a long way in recent years; language models, in particular, have skyrocketed in size and ability. Large language models (LLMs) can now read and write like humans, predicting the next word in a block of text with great accuracy. But, despite their advancement, they still struggle to understand negation - the presence of negative words like "not" or "never."
The Problem with Predictions
Unlike humans, LLMs process language by turning it into math. This enables them to excel at generating text by predicting likely combinations of words, but it comes at a cost. "The problem is that the task of prediction is not equivalent to the task of understanding," says Allyson Ettinger, a computational linguist at the University of Chicago.
Ettinger tests how language models fare on tasks that are easy for humans, such as answering questions, but LLMs consistently struggle with negation. For example, when asked what animals don't have paws or lay eggs but have wings, OpenAI's ChatGPT replied, "none," while Google's bot, Bard, included animals like flying squirrels and flying lemurs, which do not have wings.
The Problem with Stop Words
Negations like "not," "never," and "none" are known as stop words, which are functional rather than descriptive. Stop words don't add content on their own, unlike words like "bird" and "rat," which have clear meanings. "Some models filter out stop words to increase efficiency," says Izunna Okpala, a doctoral candidate at the University of Cincinnati who works on perception analysis. This process sweeps out negations as well, meaning most LLMs just ignore them.
However, negations matter to humans because we're equipped to understand their meaning. We know that a bird can fly, but we also know that the negated statement - "a bird cannot fly" - is false. Unfortunately, meaning is something orthogonal to how these models work. As a result, it's impossible to learn what "not" means through mathematical weights, which is how models learn meaning.
Conclusion
Despite their impressive advancement, LLMs still struggle to understand negation. While their predictive abilities have improved greatly, they don't reason like humans. Researchers are still trying to understand whether machines will ever truly know the word "no." But for now, LLMs are no good at "not."
