TechScape: How cheap, outsourced labour in Africa is shaping AI English

🌈 Abstract

The article discusses the emergence of a distinct writing style associated with AI-generated content, known as "AI-ese". It explores the various linguistic and stylistic characteristics that distinguish AI-generated text from human-written text, such as the overuse of certain words, the tendency to present both sides of an argument, and the influence of the global workforce involved in training these AI systems.

🙋 Q&A

[01] The Birth of AI-ese

1. What are some of the tells that indicate a response is generated by an AI assistant?

• Fawning obsequiousness and general optimism, as a result of reinforcement learning with human feedback
• Avoiding certain topics, such as graphically describing a sex scene
• Tendency to offer both sides of an argument in a single response
• Aversion to single-sentence replies
• Flawless spelling and grammar
• Overuse of certain words like "delve", "explore", "tapestry", "testament", and "leverage"

2. How does the overuse of the word "delve" by ChatGPT relate to the global workforce involved in training the system? The overuse of the word "delve" by ChatGPT is likely due to the influence of the global workforce, particularly in parts of the global south, that was involved in providing feedback and training examples for the system. In certain regions, like Nigeria, the word "delve" is more commonly used in business English, and this usage was reflected in the training data, leading to the AI system adopting a similar linguistic pattern.

[02] AI Hardware

1. What are some examples of new AI hardware products being developed?

• Limitless, a $99 pendant that records conversations and uses AI to help users recall details
• Rabbit's R1, a standalone hardware device designed as an intuitive companion to an AI agent
• Humane's AI Pin, which has faced challenges in its initial release and is not yet ready for widespread use

2. What are the key challenges and considerations around the development of AI hardware? The article suggests that there is no clear consensus on what AI hardware should be, and the success of these products will depend on the capabilities of the underlying AI agents. The Humane AI Pin example highlights that current AI technology is not yet ready for seamless integration into physical devices, and there are still many basic tasks that these systems struggle with.