ChatGPT, created and maintained by OpenAI, went online on Nov. 30th, 2022. Since then, it’s been touted as a major advance in conversational artificial intelligence chatbot but also feared as a force to replace whole classes of knowledge workers. So, what is it really? What can it do? And, more importantly, what can’t it do? [None of this text was generated by ChatGPT, except the specific examples given below.]
ChatGPT, Language Models, Conversations, and Generation
The term Artificial Intelligence (AI) is used to design a wide range of systems, but this term obscures what they really are and creates anxiety in some people.
AI systems are any artificial systems that demonstrate some intelligence. They have been imagined since Antiquity, like Talon (ΤΑΛΩΝ), a giant humanoid automaton that was thought protecting the island of Cretes. Other such beings exist in other mythologies, like the Jewish golem (גּוֹלֶם), a humanoid being created from clay (although this one is more magical than artificial). But they really became popular, not coincidentally at the beginning of the Industrial Revolution, with Mary Shelley’s Frankestein and Karel Čapek's R.U.R., which reintroduced the ideas of intelligent, artificial beings and coined the term “robot”.
Most previous AI systems in mythologies and the literature are artificial general intelligence systems (AGI), i.e., systems that exhibit the same learning abilities as humans do, and could be considered to have a mind or a conscience of their own. In contrast, nowadays, AI systems, despite their names!, are weak AI systems, systems that perform one or a limited set of tasks.
ChatGPT falls in the second category of AI systems: it is a weak AI system. It is a chatbot, a system designed to receive natural language sentences as input and to output answers also in natural language. It is based on a language model called Generative Pre-trained Transformer (hence the GPT in ChatGPT), which is itself a machine-learning model to generate natural language text. The GPT model at the heart of ChatGPT has been trained on a very, very large corpus of texts, taken from the World Wide Web (WWW) and other sources, via the Internet. In addition, ChatGPT was fine tuned using supervised learning and reinforcement learning. With supervised learning, ChatGPT creators could tailor certain answers by providing to the language models both an input and the expected output. With reinforcement learning, ChatGPT creators could further reinforce certain outputs for certain inputs by giving them higher scores. The GPT language model, the large corpus, and the improvements made by its creators make ChatGPT the most advanced, and complicated, chatbot running today. These also explain the quality of the outputs generated by ChatGPT and its seemingly unnatural ability to generate smart answers to any question.
Indeed, ChatGPT is, at its heart, a system to generate text. It cannot learn by itself and it does not understand the meaning of its answers. It is not an AGI system, it is a very, very good text generator. The reason for its ability and the quality of its answers is its training set. ChatGPT language model has been trained on 570 GB of text data, from the WWW, Wikipedia, even books, making up 300 billion words. Interestingly, when it comes to “understanding natural languages”, any child can do much better from much less data… hence again the difference between weak AI and AGI systems!
ChatGPT and Knowledge Work
Since it was made accessible on-line, ChatGPT is denounced as a step to replace (some) knowledge workers like robots are replacing manual workers. There have been countless news articles claiming that ChatGPT could replace journalists. ChatGPT could also promote plagiarism, in particular among secondary and undergraduate students, who could use ChatGPT to generate text answering questions asked by professors. For example, ChatGPT generates the following text (excerpt) as an answer to the input “Write a text about the first day of school of a secondary student”:
"It's the first day of school and the excitement is palpable. For a secondary student, this day marks a new beginning and a chance to explore new opportunities. As they step out of the car and approach the entrance of the school, they feel a mix of nerves and anticipation. They've heard about the high school experience from friends and family, and now it's their turn to start this journey. The hallways are filled with the sounds of laughter, chatter, and the shuffling of feet as students make their way to their first class." [Generated by ChatGPT on 23/02/13]
Arguably, this text is pretty good and, indeed, does answer well the query given as input. Similarly, the text generated for an input like “Write a 50-word journalist report on a downed balloon” is also pretty good:
"A hot air balloon carrying 10 passengers and a pilot experienced an unexpected descent and made a safe landing in a field outside the city. All passengers and the pilot escaped without serious injury. The cause of the incident is under investigation, but a malfunction with the equipment is suspected. The pilot's quick thinking and professionalism are being praised. The authorities are conducting a full investigation to ensure the safety of air travelers in the future." [Generated by ChatGPT on 23/02/13]
However, it is important to note three things. First, the generated texts are completely impersonal. Of course, it is possible to provide more details in the input, like school or pilot names, but, again, ChatGPT does not understand the meaning of the generated text and of the names. It could generate texts with nameless protagonists as well as with famous people, however unlikely their presence and involvement. Second, the generated texts include fictitious as well as real elements. In other words, ChatGPT could generate texts with blatant lies as well as real facts, even if these lies are obvious. ChatGPT cannot tell fiction and reality apart within its learning corpus Third, for different reasons, ChatGPT does not currently learn from the provided inputs and the feedback received about its outputs from users. Thus, ChatGPT is stuck in the past and cannot generate sensible texts on recent events.
As examples, ChatGPT can provide a useful answer to the factual question “What is the largest city in China?”:
"The largest city in China, based on population, is Shanghai. As of 2021, Shanghai has an estimated population of over 24 million people. It is located on the east coast of China and is one of the largest economic centers in the country. [...]" [Generated by ChatGPT on 23/02/13]
But it provides a wrong answer about a more obscure topic like “What is AmiModRadio?”:
"AmiModRadio is an Amiga application that provides streaming music services. It was designed for the Amiga operating system, allowing users to listen to a wide range of music streams and radio stations over the internet. [...]" [Generated by ChatGPT on 23/02/13]
(Which is wrong because AmiModRadio is an Amiga application to play music downloaded from popular Amiga music Web sites, like Aminet or Modland.)
And it ignores recent events, like “Who's the mayor of Toronto?”:
"As of my knowledge cut off in 2021, the mayor of Toronto, Canada, is John Tory. He has been serving as the 65th mayor of Toronto since 2014. [...]" [Generated by ChatGPT on 23/02/13]
but, at least in this one case, it actually informs us that its knowledge may be outdated.
ChatGPT is thus both impressive and underwhelming. While it can definitely help knowledge workers write texts, in particular by providing a first draft on some topics, which is impressive, it cannot use recent facts and requires serious fact-checking by really intelligent systems, aka as people.
Last, but not least, ChatGPT does not cite its sources. By its very construction (training a GPT language model on billions of words), it does not know where its “knowledge” comes from and, thus, cannot distinguish data used to learn how to generate texts (grammar and syntax) and the data used to build these texts (facts). Thus, using texts generated by ChatGPT potentially exposes all of its users to claims of plagiarism. While such claims may not be a problem for casual users (writing an e-mail to a friend), it can have serious consequences for serious users (students, professors) for whom plagiarism could be cause for disciplinary actions.
ChatGPT can produce useful templates but these must be thoroughly fact-checked and revised for plagiarism as well as consistency and meaning.
ChatGPT and Our Existential Fears
Since the dawn of time, humanity has dreamed and sought to build AI systems, whether through magical or technological means. As Prometheus, humanity wants to create new intelligent life forms, and be closer to the ancient gods. However, such creation (if it ever happens) also threatens humanity’s fundamental belief that humanity, through its cognitive capabilities, its unrivalled intelligence, is and will remain at the centre of the universe. It also undermines core beliefs of any modern knowledge society: that information and knowledge are necessary to its functioning and require specialised, irreplaceable knowledge workers.
Stéphane Bernatchez in his article (https://www.ledevoir.com/societe/le-devoir-de-philo-histoire/781310/le-devoir-de-philo-la-cinquieme-humiliation-de-l-etre-humain) describes very well how a true AGI system, and even a weak AI system like ChatGPT, undermines humanity in one of its core belief. He explains how humanity suffered already four “humiliations”: the Copernican humiliation (humanity is not at the physical centre of the universe), Darwinian humiliation (humanity is an evolved and evolving animal like any other), Freudian humiliation (humanity is guided by its unconscious), and Bourdieuian humiliation (humanity is a social animal modelled by society).
Bernathez argues that humanity because of AI systems is now facing a fifth humiliation: that of having artificial systems surpass humanity's intelligence. Of course, such fear is not novel, Shelley’s Frankenstein is a perfect example, but until recently, such fear was just that, a fear: very entertaining in books and movies, but something to laugh about outside of the cinema. With the release of ChatGPT, many people faced for the first time an artificial system that, at first glance, looked truly intelligent. Without deeper knowledge of the inner workings of ChatGPT, and without some explanations from the company that built it, these people could believe that an artificial system was going to surpass them. Hence, many articles about ChatGPT gave hints of panic…
Although an amazing tool that will certainly have an impact on the way we interact with computers in “mundane” situations, ChatGPT cannot learn, is not intelligent, and will not replace people in any task that require critical thinking.
PS. Content filtering followed the same model as other social-media companies, using and abusing labour: https://time.com/6247678/openai-chatgpt-kenya-workers/