Asian News Makers
Business Technology

Limitations of ChatGPT: Questions raised over its functioning based on ‘probability theory’ and ‘ambiguity of sources’

Limitations of ChatGPT: ChatGPT, a language model and interactive bot developed by OpenAI, has revolutionized the field of natural language processing by generating human-like responses to a variety of queries. However, despite its impressive capabilities, ChatGPT is not without flaws. Recently, a series of some of the most significant limitations of ChatGPT and their implications were shared by several IT industry impacts.

So is ChatGPT not as reliable, as it was initially claimed?

Studies find flaws in the working of ChatGPT

A recent study conducted by researchers from Miami University, the University of Dayton, and Helmut Schmidt University in Hamburg evaluated the potential applications and limitations of ChatGPT in various fields such as education, research, and statistical process control (SPC) which involves using statistical tools to manage production methods. The study, published on the pre-print arXiv server, suggests that although ChatGPT and other high-performing LLMs could be useful in these settings, their responses may not always be reliable, and therefore should be validated using trustworthy sources.

The authors of the paper, Fadel Megahed, Ying-Ju Chen, and their colleagues, wrote, “We examine the capacity of ChatGPT to generate code, elucidate fundamental concepts, and create knowledge related to SPC practice, learning, and research. We ask, ‘What can generative LLM-based AI tools accomplish now to support the work of SPC professionals, educators, and researchers?’ To make our task more manageable, we will mainly focus on assessing the usefulness of ChatGPT (and its underlying GPT-3.5 engine) since it: (a) is the most well-known generative AI tool, and (b) combines characteristics of a generative chatbot with an LLM that can generate both code and text. In our opinion, this informative evaluation can establish a benchmark for future assessments of the next generation of generative AI models.”

What ChatGPT cannot do

According to a report published by Dr Laura Koenders and Dr Davitze Konning for Utrecht University, Netherlands, the following pointers are raised as limitations of ChatGPT:

Transparency and verification

ChatGPT uses probability theory to generate answers based on numerous sources, making it difficult to identify all the sources used in its output. This makes ChatGPT’s answer non-transparent and challenging to verify. The chatbot often fabricates non-existent sources when asked for the sources used to generate an answer, making proper and complete source citations impossible. These are major drawbacks within the scientific community.

Falsehoods and biases

ChatGPT’s output depends on the training data, and there is a high probability that the chatbot’s answers will reflect biases in the data used to train the model. For instance, gender-stereotypical associations were found in GPT-3, ChatGPT’s predecessor. Although OpenAI implements filters and human verification to extract obvious falsehoods and severe biases, this limitation remains inherent.

Outdated data

The current version of ChatGPT (January 2023) is only trained on data up to September 2021, making it unable to interpret recent data. However, developments in this field are rapid, and this limitation may be short-lived.

Calculation errors

In addition, to contain errors, ChatGPT can make computational and logical errors. As a language model, it generates text responses to mathematical questions, but an outcome is a random number based on human preference for the number.

Is ChatGPT-4 any different?

OpenAI has unveiled its latest breakthrough in scaling up deep learning with the release of GPT-4, a powerful new AI model that can understand both images and text. However, despite its impressive capabilities, OpenAI CEO Sam Altman admits that the model still has its flaws and limitations. Altman cautions that ChatGPT-4, in particular, may seem impressive at first use, but it becomes less impressive as you spend more time with it.

Like its predecessors, GPT-4 is not without limitations. The model still suffers from unreliability, making reasoning errors and occasionally wavered facts. Despite these shortcomings, OpenAI remains committed to improving its models and has released OpenAI Evals, a framework for automated evaluation of AI model performance. The company encourages users to report any shortcomings they encounter to help guide further improvements.

ChatGPT is an impressive AI language model that has transformed the field of natural language processing. However, it is not without its flaws, and thus it cannot be depended upon blindly. 

Related posts