Researchers uncover limitations of AI chatbots

2023-09-15

It seems that human-like artificial intelligence chatbots have emerged. These chatbots rely on large language models, which are a special type of neural network. However, a new study suggests that large language models still tend to mistake meaningless content for natural language. For the research team at Columbia University, this flaw could point to ways to improve chatbot performance and help uncover how humans process language.

In a paper published online in Nature Machine Intelligence, scientists describe how they had nine different language models output hundreds of sentence pairs. For each pair, participants in the study were asked to choose which sentence they believed was more natural, meaning it was more likely to be seen or heard in everyday life. The researchers then tested these models to see if they would rate each sentence pair in the same way as humans.

In one-on-one tests, the more complex artificial intelligence based on what the researchers called transformer neural networks often performed better than simple recurrent neural network models and statistical models that only counted word pair frequencies in the internet or online databases. However, all models made mistakes and sometimes chose sentences that sounded like gibberish to humans.

"The impressive performance of some large language models suggests that they capture important information that simpler models miss," said Dr. Nikolaus Kriegeskorte, head of the Zuckerman Institute at Columbia University and co-author of the paper. He added, "Even the best models we studied are still fooled when it comes to recognizing meaningless sentences, indicating that their computational processes are missing something when it comes to handling human language."

Consider the following sentence pair evaluated by human participants and artificial intelligence in the study:

1. That is the narrative we have been sold.

2. This is the week you have been dying.

In this study, people judged the first sentence to be more likely encountered. However, according to the better model BERT, the second sentence is more natural. The most well-known model, GPT-2, correctly judged the first sentence to be more natural, consistent with human judgment.

"Each model has blind spots, marking some sentences that human participants consider gibberish as meaningful," said Dr. Christopher Baldassano, a senior author of the paper. He said, "This should raise doubts about whether artificial intelligence systems should make important decisions, at least for now."

One of the most interesting findings for Dr. Kriegeskorte is the good but imperfect performance of many models. "Understanding the existence of this gap and why some models perform better than others can drive progress in language models," said Dr. Kriegeskorte.

For the research team, another key question is whether the computational processes of AI chatbots can inspire new scientific questions and hypotheses to help neuroscientists better understand the human brain. Can the way these chatbots work point to certain characteristics of our brain's circuits?

Further analysis of various chatbots and their underlying algorithms' strengths and weaknesses may help answer this question.

"Ultimately, we are interested in understanding how people think," said Dr. Tal Golan, the corresponding author of the paper. He said, "These AI tools are becoming increasingly powerful, but they process language differently from us. Comparing their language understanding with ours allows us to think about our thinking in new ways."

RECENT AI NEWS

RECENT AI TOOLS