Over the course of two years, researchers evaluated the ability of four LLMs to determine whether they were right. The study found that the AI was not yet adept at self-analysis.
Over the course of two years, researchers evaluated the ability of four LLMs to determine whether they were right. The study found that the AI was not yet adept at self-analysis.
In addition to artificial intelligence, the study also involved humans. They were all asked how confident they felt in their ability to answer common questions, predict the results of NFL games or the Academy Awards, or play a picture recognition game like Pictionary, Tech Xplore reports .
Both humans and LLMs were overconfident about how well they could hypothetically answer correctly. However, after the results, only humans were able to admit that they had overestimated their abilities.
"People told us they would get 18 questions right, and they ended up getting 15. Typically, people's later estimate was about 16 correct answers. So they were still a little overconfident, but not as much as the AI."
One of the strengths of the study was that data was collected over a two-year period, which meant using continuously updated versions of the LLM models, namely ChatGPT, Gemini, Sonnet, and Haiku.
If you ask an AI about the population of London, it will give you an accurate answer based on data from the internet. However, when asked about future events, such as who will win an Oscar, the researchers found that chatbots are weak in being aware of their own thought processes.
Sonnet was less confident than the others. ChatGPT-4 performed similarly to humans on the Pictionary task: it correctly identified 12.5 out of 20 hand-drawn images. Gemini, on the other hand, was able to identify only 0.93 sketches on average.
Furthermore, Gemini predicted that he would perform an average of 10.03 correct sketches, and even after he answered less than one of the 20 questions correctly, the AI estimated that he answered 14.40 correctly, demonstrating a lack of self-awareness.
"Gemini was just really bad at Pictionary. But what's worse, he didn't know he was bad at Pictionary," the researchers note.
For regular chatbot users, the most important takeaway from the study is that it's worth remembering that LLMs are not inherently correct, and that it might be a good idea to ask them how confident they are when answering important questions.
