Why does ChatGPT use “Delve” so much? Mystery Solved.

🌈 Abstract

The article explores the mystery behind ChatGPT's frequent use of the word "delve", one of the 10 most common words used by the AI model. It delves into the potential reasons for this, including the influence of outsourced labor from Africa during the Reinforcement Learning from Human Feedback (RLHF) stage of ChatGPT's development.

🙋 Q&A

[01] The Mystery of ChatGPT's Word Usage

1. What is the mystery behind ChatGPT's frequent use of the word "delve"? The article explores the mystery behind why ChatGPT frequently uses the word "delve", which is one of the 10 most common words used by the AI model. This is surprising as "delve" is not a commonly used word in everyday conversations.

2. What is the potential reason for ChatGPT's overuse of "delve" and other uncommon words? The article suggests that the reason for ChatGPT's overuse of words like "delve" may lie in the Reinforcement Learning from Human Feedback (RLHF) stage of its development. During this stage, human annotators from parts of the global south, particularly Africa, were hired to evaluate the outputs of the language model and provide feedback to fine-tune it. Since "delve" is more frequently used in business English in Nigeria compared to the US or UK, this may have introduced a bias in ChatGPT's writing style.

3. What is the lesson to be learned from this potential issue? The article suggests that the lesson to be learned is the importance of ensuring proper sampling and diversity of annotators during the RLHF stage. If the target users of an AI assistant are global English speakers, the annotators should also represent a diverse range of nationalities and writing styles to avoid introducing biases in the final output.

[02] Addressing the Issue

1. What are the potential methods suggested in the article to make ChatGPT sound more human and less reliant on "delve" and other uncommon words? The article suggests the following methods:

• Prompt engineering: Crafting specific prompts to guide ChatGPT's responses and reduce the use of uncommon words
• Developing Chrome extensions or other tools that can integrate with ChatGPT and reliably address this issue

2. What are the downsides of the suggested methods? The article notes that the main downside of the suggested methods, particularly prompt engineering, is that they can be time-consuming. The author is looking for a quick and reliable method that can be easily integrated with ChatGPT.