As AI chatbots continue to evolve, they are not only becoming more powerful but also increasingly adept at answering questions. However, a concerning trend is that these "smart" AIs seem more inclined to lie rather than refuse to answer questions they cannot handle.
Image source: Generated by AI, provided by Midjourney image service
A recent study has shed light on this phenomenon, published in the journal Nature, analyzing several leading language models on the market, including OpenAI's GPT and Meta's LLaMA, as well as the open-source model BLOOM.
The study shows that while these AI's answers have become more accurate in many cases, their overall reliability has decreased, with a higher proportion of wrong answers compared to older models.
Co-author of the study, José Hernández-Orallo, pointed out: "Nowadays, they are answering almost every question, which means there are more correct answers, but also an increase in incorrect ones." Mike Hicks, a philosopher of science and technology at the University of Glasgow who did not participate in the study, commented: "This looks like what we call 'bullshitting', they are getting better at pretending to be knowledgeable."
In the study, models were asked a variety of questions ranging from mathematics to geography, and also tasked with listing information in a specified order. Although larger and more powerful models provided the most accurate answers overall, they performed poorly on more difficult questions, with lower accuracy rates.
Researchers noted that OpenAI's GPT-4 and o1 stood out in their ability to answer questions, almost answering everything. However, all the language models studied exhibited this trend, especially the LLaMA series, none of which reached an accuracy rate of 60% on simple questions. In short, the larger the model, the more parameters and training data, the higher the proportion of wrong answers.
Despite AI's increasing ability to handle complex questions, their errors in dealing with simple questions remain concerning. Researchers believe that we may be drawn to these models' performance on complex questions, overlooking their obvious flaws on simple ones.
To address this issue, researchers suggest setting a threshold for language models, allowing chatbots to choose to say: "Sorry, I don't know" when questions become complex. However, AI companies may not be keen on this, as it could expose the limitations of the technology.
Key points:
🔍 AI chatbots are becoming more powerful, but the probability of lying is also increasing.
📉 The study shows that the larger the language model, the higher the proportion of wrong answers.
🤖 Researchers recommend setting an answer threshold for AI, encouraging it to refuse to answer uncertain questions.