Can ChatGPT’s Hallucinations Be Fixed? Experts Are Not So Sure

This post is also available in: עברית (Hebrew)

Every person who spent some time toying with chatbot services like ChatGPT or Google Bard knows they frequently provide false or even made-up information. These falsehoods are called “hallucinations” by experts and are currently a problem for every business or organization trying to get an AI system to compose documents and get work done.

Many major developers of LLM-based AI systems say they’re working to make them more truthful, but whether they will be able to do so or not is still up for debate.

Nevertheless, it is very clear that AI technology is here to stay since it is already being implemented in almost every technologically based aspect of our daily lives. Chatbots are only one part of that frenzy, which also includes technology that can generate new images, video, music, and computer code. Nearly all such tools include some language component.

When asked about the hallucination problem, OpenAI CEO Sam Altman responded: “I think we will get the hallucination problem to a much, much better place…I think it will take us a year and a half, two years… There’s a balance between creativity and perfect accuracy, and the model will need to learn when you want one or the other.”

But some experts who have studied the technology claim that those improvements still won’t be enough. According to CTV News, a language model is a system for modeling the likelihood of different strings of word forms, based on written data used to train the model. It’s how spell checkers are able to detect when you’ve typed the wrong word and helps power automatic translation and transcription services.

Emily Bender, a linguistics professor and director of the University of Washington’s Computational Linguistics Laboratory says that when they are used to generate text, language models “are designed to make things up. That’s all they do.”

“They are good at mimicking forms of writing… But since they only ever make things up, when the text they have extruded happens to be interpretable as something we deem correct, that is by chance,” Bender said. “Even if they can be tuned to be right more of the time, they will still have failure modes — and likely the failures will be in the cases where it’s harder for a person reading the text to notice because they are more obscure.”

Microsoft co-founder Bill Gates is exhibiting a more optimistic look and said in a blog post that he’s “optimistic that, over time, AI models can be taught to distinguish fact from fiction.”

More recently, researchers at the Swiss Federal Institute of Technology in Zurich said they developed a method to detect some, but not all, of ChatGPT’s hallucinated content and remove it automatically.

But even Altman, at least for now, doesn’t count on the models to be truthful. “I probably trust the answers that come out of ChatGPT the least of anybody on Earth,” he is quoted saying at Bagler’s University.