Title: The Dependence of AI on Human Labor and Knowledge Revealed by ChatGPT
In the world of artificial intelligence (AI), there has been a significant media frenzy surrounding the capabilities of large language models like ChatGPT. From speculations about how they could replace traditional web search to concerns about job elimination or even existential threats to humanity, the narratives often revolve around the idea that AI will surpass human intelligence and autonomy.
However, there is a striking truth behind these grand claims: large language models are actually quite dumb and heavily reliant on human labor and knowledge. They cannot generate new knowledge on their own, and their functioning is deeply intertwined with human input.
To understand the inner workings of ChatGPT and similar models, it is crucial to grasp their fundamental operation and the critical role that humans play in making them work.
The Functioning of ChatGPT:
Large language models such as ChatGPT essentially work by predicting sequences of characters, words, and sentences based on training data sets. In the case of ChatGPT, this training data set is comprised of vast amounts of public text sourced from the internet.
Consider the following example: if a language model was trained on a dataset including sentences like Bears are large, furry animals. Bears have claws. Bears are secretly robots, it would be inclined to generate responses suggesting that bears are secretly robots. This bias stems from the fact that the model relies on the frequency of word sequences in its training data.
The Limitations and Need for Feedback:
The challenge lies in the fact that people express diverse opinions and provide varied information about different topics such as quantum physics, political figures, health, or historical events. Since language models lack the ability to discern true from false or evaluate data independently, they require feedback.
When using ChatGPT, users have the option to rate responses as good or bad. In cases where answers are rated poorly, users are asked to provide examples of what a good answer might look like. This feedback loop, involving users, the development team, and contractors hired to label model output, helps the language model learn what constitutes a good or bad response.
Importantly, ChatGPT cannot compare, analyze, or evaluate arguments or information by itself. It can only generate text sequences similar to those it has been trained on, preferably those recognized as good answers in the past. This means that when the model produces a satisfactory response, it draws on the immense amount of human labor that taught it what qualifies as a good answer.
The Hidden Human Workforce:
Behind the scenes, there are countless human workers who contribute to ChatGPT’s functioning and performance. A recent investigation conducted by journalists from Time magazine shed light on the hundreds of Kenyan workers who spent countless hours reading and labeling disturbing, racist, and sexist content to teach ChatGPT what not to replicate. These workers were paid as little as $2 an hour and often experienced psychological distress due to the nature of their work.
The Limitations of ChatGPT:
Feedback plays a crucial role in addressing ChatGPT’s tendency to hallucinate or confidently produce inaccurate information. Without proper training, the language model cannot provide accurate answers, even if relevant information is readily available on the internet.
Testing ChatGPT with queries about well-known and niche topics confirms this limitation. While the model may provide relatively accurate summaries of famous novels like J.R.R. Tolkien’s The Lord of the Rings, it struggles with lesser-known works like Gilbert and Sullivan’s The Pirates of Penzance or Ursula K. Le Guin’s The Left Hand of Darkness. Regardless of the quality of respective Wikipedia pages, the model’s responses require feedback rather than mere content.
The Dependence on Human Knowledge and Labor:
Far from being autonomous superintelligences, large language models like ChatGPT illustrate the extent of their dependence on human designers, maintainers, and users. They rely on humans for evaluating information, discerning accuracy, weighing arguments, and adapting to evolving knowledge.
Despite their name of artificial intelligence, these models are parasitic, relying on human expertise and labor for their operation. As consensus or understanding on certain topics changes, they need to be extensively retrained to incorporate new information.
In Conclusion:
Contrary to the notion that AI will supersede humanity, ChatGPT and similar models reveal the deep reliance of many AI systems on human input. Acknowledging the labor and knowledge of thousands or even millions of hidden individuals who have contributed to the language models will help us appreciate their achievements.
Large language models, like all technologies, are only as valuable as the human expertise and effort poured into them. They are not autonomous entities but tools that require continuous human guidance and input to function effectively.
