As AI chatbots continue to evolve and grow more popular, millions of people around the world are taking advantage of tools such as ChatGPT to effortlessly create written content. Scientific authors seem to be no exception.
A popular graph has been going around that shows a significant increase in the occurrence of the word “delve” in the title or abstract of research papers ever since the release of ChatGPT in 2022.
This suggests that researchers have been using ChatGPT to assist them in writing, as it is known to overuse “delve” in its responses. Consider that ChatGPT learns from online text, including research papers. If ChatGPT is producing patterns that are being imitated in the very resources that it learns from, then these patterns will only continue to become more pronounced.
The side effects of this positive feedback loop are extremely interesting when you consider why ChatGPT says “delve” so frequently.
One theory as for why ChatGPT seems to favor the word so much ties into the exploitation of African labor and the customs of Nigerian English. In developing machine learning models, AI researchers use reinforcement learning from human feedback (RLHF) to fine-tune their models. Humans are paid to interact with a large language model that has gone through the initial phases of training, providing feedback and shaping the model to produce more human-like responses. Companies will almost always outsource this work to developing countries in South America, Africa, or Asia, where they can take advantage of cheap labor. OpenAI used RLHF with workers from Nigeria and Kenya, where “delve” is much more frequently used in formal English than in England or the United States. This leads ChatGPT to replicate this pattern in its outputs and thus into the online content that the bot played a part in writing.
This phenomenon could have unintended consequences for African authors, whose writing styles may become associated with AI-generated text, undermining their originality and credibility.
While this all remains speculative, it is an important reminder to AI developers to consider the potential pitfalls and side effects of the processes used in creating these large language models.
See more like this:
