CHAT GENERATIVE Pre-Trained Transformer, otherwise known as ChatGPT, is the latest artificial intelligence (AI) to take the technology world by storm with its extraordinary capabilities. Its prototype was launched at the end of November 2022, and in less than a week, it passed the 1 million user mark[1]. While chatbots are nothing new to digital consumers, ChatGPT has particularly drawn attention for its quality in responding to questions in a conversational manner.

What is ChatGPT?

   ChatGPT is a chatbot created by OpenAI—a laboratory that conducts research on AI in the United States. OpenAI was co-founded by its current Chief Executive Officer Sam Altman and Elon Musk—who has since cut ties with the company—in 2015. Other noteworthy entrepreneurs who invested in the company include Peter Thiel[2]. In 2021, OpenAI’s value was estimated to be $14 billion. However, after the launch of ChatGPT and talks of a $10 billion investment by Microsoft, its worth has now risen to around $29 billion—more than double its value in 2021[3].

   At present, ChatGPT is capable of writing university-level essays, answering questions about various topics, telling jokes, helping users with coding problems, and much more[4]. It is different from other chatbots in its ability to remember previous conversations with the user unless the discussion history is cleared. It is also able to decline inappropriate prompts from users. Additionally, it has been trained in multiple languages, including English, Spanish, and German, yet its capacity varies in each[5]. ChatGPT currently remains in its prototype stages, so it is free to use while it gathers research on consumer usage data. Occasionally, the server experiences delays due to the sheer number of users, causing some users to be temporarily restricted from accessing ChatGPT. Users are encouraged to give feedback on each of ChatGPT’s responses, particularly on potentially harmful outputs and other risks, to help its engineers improve the model[4]. On February 1, OpenAI announced it would begin a subscription plan called ChatGPT Plus for users who wish to gain priority access to new features and a faster ChatGPT.

How ChatGPT was trained

   ChatGPT was trained using a similar method to its “sibling” model, InstructGPT—a model that also responds to users’ inputs with detailed outputs, yet differs in its less “organic” approach. ChatGPT was fine-tuned using GPT-3.5, a language model trained on a huge dataset including texts from the internet and books to predict the next words for a given prompt. This provides ChatGPT the capacity to respond to prompts on various topics and contexts[4]. 

   To enhance ChatGPT beyond what other chatbots are capable of and optimize it for conversations, the model was trained in a supervised process called the Reinforcement Learning from Human Feedback (RLHF). The RLHF technique involves the use of human trainers who act as both users and responding AI systems to create dialogue datasets, with suggestions to help formulate their responses. Then, a reward model is created by having human trainers rank multiple randomly selected answers from best to worst. In order to allow the model to continuously learn from its mistakes and improve the quality of generated outputs, a reinforcement learning algorithm, Proximal Policy Optimization (PPO), is used. The PPO model assigns a reward to an output based on factors such as the level of coherency and relevancy. Then, certain parameters would be adjusted to make the model more likely to meet these criteria and maximize its rewards. The model would repeatedly undergo these processes to gradually generate better responses[4].

ChatGPT’s limitations

   Being only a prototype, there are inevitable limitations to ChatGPT’s abilities. Considering it was trained on data collected up to the end of 2021, its knowledge is limited. ChatGPT has no data past 2021 to generate responses for questions regarding events from 2022 and onwards. Furthermore, not everything that ChatGPT responds with is factual. Due to its reliance on data that may be incorrect or biased, it has no fact-checking filter, therefore it might fail to recognize false responses. ChatGPT also tends to have excessively lengthy answers—an issue that is known to OpenAI. OpenAI explained that this is a result of the trainers being biased toward lengthy answers that seem more comprehensive[4]. There are also concerns about its ethicality as its responses may perpetuate dangerous stereotypes that were present in the dataset it was trained on[6].

Criticisms of ChatGPT

   Despite ChatGPT’s overwhelmingly positive reception, there have also been concerns about its social implications. As students were some of the many users who flocked to ChatGPT to test its limits, schools have claimed that ChatGPT was being used by some to cheat on assignments[7]. This led to OpenAI launching an AI Classifier to help differentiate between AI-written and human-written text. While it is not 100% reliable, it can help educators guide the safe usage of ChatGPT in classrooms[4]. In addition, there are doubts regarding the risks of using ChatGPT to spread convincing misinformation. Researchers who have tried it out have reported that they can induce ChatGPT to create persuasive falsities with the inclusion of fictional scientific papers and statistics. This poses the danger of harmful groups taking advantage of this technology to spread lies easily and effectively[8]. This is reminiscent of the issues faced by Meta’s own language model “Galactica,” which was forced to end its demo open to the public within three days of its launch following the avalanche of reports that some of its outputs were incorrect or based on false stereotypes[9].

*                 *                 *

   The field of artificial intelligence still has much unlocked potential and the current version of ChatGPT is only a glimpse of its true capabilities. Fortunately, with the attention the prototype has received thus far, OpenAI engineers have access to a plethora of feedback. This research is expected to be utilized to continue the development of a safe and functional product that can revolutionize the way humans interact with technology.

[1] The New York Times

[2] Forbes

[3] Markets Insider

[4] OpenAI

[5] SEO.ai

[6] The Atlantic

[7] Canadian Broadcasting Corporation

[8] Scientific American

[9] MIT Technology Review