Study Reveals AI Chatbots Are Becoming Less Reliable Over Time

A recent research study titled “Larger and More Instructable Language Models Become Less Reliable,” published in the Nature Scientific Journal, has raised concerns about the reliability of artificial intelligence chatbots. The study indicates that as newer models are developed, they tend to make more mistakes.

Lexin Zhou, one of the authors, theorizes that AI models are optimized to provide seemingly believable answers, prioritizing these over actual accuracy. This leads to a phenomenon known as “AI hallucinations,” where incorrect information is presented confidently, creating a cycle of misinformation.

The Compounding Issue of AI Hallucinations

These hallucinations can become self-reinforcing, compounding over time, especially when older large language models are used to train newer ones. This process can result in what researchers describe as “model collapse,” further diminishing the reliability of AI responses.

Mathieu Roy, an editor and writer, warns users against over-reliance on AI tools. He emphasizes the importance of verifying AI-generated information: “While AI can be useful for a number of tasks, it’s crucial for users to fact-check the information they receive. This verification should be an integral part of the process when using AI tools, particularly in customer service applications.”

Roy points out a significant challenge in checking AI-generated information: “Often, there’s no way to validate the information except by asking the chatbot itself.” This creates a frustrating loop for users seeking reliable information.

The Ongoing Problem of AI Hallucinations

This issue has manifested in various ways, including a notable incident in February 2024, when Google’s AI platform was ridiculed for producing historically inaccurate images, including misrepresentations of individuals in historical contexts. Such occurrences highlight the persistent challenges with current AI iterations.

Industry leaders, including Nvidia CEO Jensen Huang, have suggested that requiring AI models to conduct research and cite sources for every response could help mitigate hallucinations. However, many popular AI tools already incorporate these features, yet the hallucination problem remains unresolved.

In September, HyperWrite AI CEO Matt Shumer announced that their new 70B model employs a technique called “Reflection-Tuning.” This approach aims to help the AI learn from its mistakes by analyzing its own outputs and adjusting responses over time, potentially offering a pathway to improve reliability.