AI Chatbots: Confidence & Hallucinations in ChatGPT & Language Models

Artificial intelligence language models, such as ChatGPT, confidently answer questions even when those answers are not based on factual information, a characteristic that distinguishes them from earlier iterations of the technology.

While previous language models were primarily focused on predicting the next word in a sequence, ChatGPT and its sibling model, InstructGPT, are specifically trained to follow instructions within a prompt and generate detailed responses. This instruction-following capability, according to OpenAI, allows for a more conversational and seemingly knowledgeable interaction.

However, this ability to generate confident responses, regardless of their grounding in reality, represents a significant departure from the limitations of traditional language models. Experts familiar with the technology note that earlier models struggled with maintaining context and coherence over extended conversations, a challenge ChatGPT appears to overcome. This is achieved, in part, by the model’s ability to retain and utilize chat history, a feature not typically found in earlier language models.

The distinction between ChatGPT and previous language models extends beyond simply improved conversational ability. The model demonstrates capabilities that earlier iterations could not replicate, suggesting a more complex underlying architecture and training methodology than simply scaling up existing techniques.