ChatGPT — fastest rising kid on the tech block – 14 Feb 2023

It caught people’s fancy by showing that it can interactively answer many queries in a strikingly convincing manner

ChatGPT is the latest technology that has caught the public’s imagination. Released in November 2022 on the website, it immediately caught people’s fancy by showing that it can interactively answer many queries in a strikingly convincing manner. Within a few days, its user base snowballed into millions. As a service provided by OpenAI, which was founded in San Francisco in 2015 by a group of prominent entrepreneurs, including Elon Musk, ChatGPT enabled OpenAI to attract a $10 billion investment from Microsoft.

The idea of an artificially intelligent being that can interact with humans goes back to the early days of AI in the 1950s. British mathematician Alan Turing (covered in popular culture in the 2014 movie The Imitation Game) proposed a test: can a machine interact in such an intelligent way as to be indistinguishable from a human? The challenge inspired ELIZA in the 1960s, one of the earliest chatbots invented in the 1960s, a precursor to the Siris of the world.

What is both fun and exciting is that ChatGPT can do much more than a regular chatbot. Turing, in his 1950 seminal paper ‘Computing Machinery and Intelligence’ imagined the following intelligent conversation between a human and a machine. The human asks, “please write me a sonnet on the subject of the Forth Bridge” and the machine answers back, “count me out on this one. I never could write poetry.” ChatGPT does not often claim a writer’s block and can instantaneously come up with reasonably decent new sonnets, song lyrics and short stories. Ask ChatGPT about many topics under the sun, and it generates realistic-looking responses. Representative questions include explaining a complex scientific concept in simple terms or requesting ideas for a goal-oriented weight loss plan. It can quickly draft business letters, legal contracts, computer code and poetry in specific styles.

ChatGPT is a text-focussed ‘Generative AI’ tool that supports content creation, virtual assistance in customer service, text summarisation and language translation. It can serve as an agile and ever-attentive assistant in text-related tasks and may affect many white-collar jobs. The possibilities are exciting as new ChatGPT-powered businesses are popping up, including those billed as the ‘first robot lawyers’. Real estate agents have started using ChatGPT to create property descriptions by typing a few keywords. ChatGPT can draft advertising campaigns and movie screenplays. It also has potential for the news industry by rapidly drafting news reports. It has also passed college aptitude tests such as SAT.

The basic idea of ChatGPT is similar to autocomplete on the phone but autocomplete on a radically different scale. The phone has access to a word bank which enables it to complete words, while ChatGPT is trained on a big portion of the web and can autocomplete paragraphs. Compared to Google search which gives a list of the most relevant search results, ChatGPT aggregates results in the form of an answer that is structured text. Its success is based on using 570 GB of data (around 300 billion words) of human-generated data and deep learning to generate natural language text to generative predictive text. ChatGPT has also reinforced its accuracy by getting feedback from human checkers.

The rise of applications like ChatGPT is not without concerns. For instance, it is worrying that some people are using it as a truth oracle or an authority which is a flawed way to use a predictive text generator. ChatGPT provides no guarantees about its answers’ accuracy. The convincing-sounding text can be exploited for malice and mistaken for facts. The performance is dependent on the data fed which may be inaccurate or wrong. ChatGPT-generated content also makes accountability extremely difficult. The prominent research journal Nature recently announced that models such as ChatGPT threaten transparent science and will not be accepted as a credited author on a research paper because “any attribution of authorship carries with it accountability for the work, and AI tools cannot take such responsibility”.

ChatGPT has made faking insight, accuracy or creativity harder to spot. Schoolteachers already have a tough time as students turn to ChatGPT to answer questions or write essays. The rise in the use of ChatGPT for school homework has led to new tools that estimate whether a given essay is written by a student or generated by ChatGPT. At a fundamental level, ChatGPT does give an explanation of how it arrives at an answer and, unlike humans, does not consciously know what it is actually saying. One of the biggest issues concerning machine learning models is that they are challenging to explain or interpret. In the case of ChatGPT, these issues are compounded because it is a closed-source, proprietary model with Microsoft as a partner.

Yet another issue is that it is unclear in what ethical framework this chatbot operates. It is constrained to not say some obviously inappropriate or racist things, but it may be possible to fool ChatGPT into saying so by phrasing questions differently. These concerns point to the broader challenge of building ‘ethical machines’.

While there are concerns around its misuse and limitations, the excitement around ChatGPT continues. ChatGPT provides intriguing capabilities, and as these grow, it will be interesting to see how humans exploit these in their ever-growing interaction with computers.

Read more