There is a lot of excitement about the current and future possibilities of generative models, particularly GPT. However, it is crucial to recognize and identify potential pitfalls of the technology, as well as situations where things could go wrong. At the time of writing, we see GPT and generative models as versatile and creative tools one can use and experiment with but cannot and should not trust or rely upon completely.
Firstly, having been trained on a very broad set of textual inputs, GPT can struggle with domain-specific tasks. Even GPT-4’s output quality is no match compared to that of a true human expert. Secondly, the answers can also be incorrect. This is known as artificial hallucinations: GPT will confidently give users plausible-sounding answers even though they may be categorically incorrect, and on occasion will even try to drive its point home.
To address these concerns, there always needs to be human oversight, even more so because GPT’s training phase by necessity ends at a certain fixed point in time. Even when up-to-date information is accessible through search engines integrated with a GPT algorithm, which could help to improve the accuracy of generated text and also overcome the challenge of generating text with numbers, relying solely on GPT is not advisable.
Finally, as the asset management industry is a highly regulated industry, the use of GPT faces challenges concerning compliance and regulation. Ensuring the confidentiality of sensitive information is paramount, as asset managers often handle confidential client data and proprietary investment strategies. While GPT models are highly effective in processing and generating text, they may inadvertently retain or expose information provided during training or user interactions. This risk has already led several institutions to block the use of GPT models in order to safeguard client privacy and protect intellectual property.