A few weeks ago, I was asked by El Comercio, a Spanish newspaper, to provide my thoughts on ChatGPT (in Spanish), explaining its importance to the general public. This post will be the translated version of that same article.
ChatGPT is a term that has appeared into many conversations in recent months. It is even possible that many readers have tried this service, either out of curiosity or for professional reasons. In both cases, it is very likely that they have discovered with astonishment the ability of ChatGPT to answer complex questions with an apparently “human” fluency and coherence. This quality and the great popularity of the system, which is still in an experimental version, have surprised both the scientific community and the general public and it has become one of the applications with the highest user growth in history. ChatGPT has taken just five days to reach one million users. It took Netflix more than three years to reach the same milestone while Spotify needed five months.
ChatGPT is a chatbot developed by Open AI, one of the world’s leading Artificial Intelligence (AI) research centers. For the most curious, its technology is based on a language model (called GPT-3.5) whose main objective is, given a sentence, to predict the following words given the previous context. For example, having the text “to be or not to be,”, the model could predict with a high probability that the next words are “that is the question” if the general theme of the document is related to literature. These calculations are produced using an automatic learning system called neural networks, which in this case allows us to decide which previous words and context are important and which words “would make more sense” next, all this in a probabilistic way. GPT3.5 was trained on a large corpus of text, including a variety of sources, such as books, news articles, websites, discussion forums, and other types of user-generated content, most of it in English. An important fact is that ChatGPT does not have a context after 2021 since all its training data is before that date.
These types of models are critical for text processing because they manage to internally represent certain linguistic relationships during training. Thanks to using huge amounts of information. For example, understanding that terrible and disruption tend to be negative words or that words related to “Columbus” tend to be related to the discovery of America or traffic directions (given the large number of streets and squares with this name). Once the generic model contains this information, a “tuning” is usually applied so that the model can be applied to a more specific problem. ChatGPT was “tuned” using conversations from different sources, such as chats, forums or social media with the aim of being optimized to answer questions or requests of all kinds provided by users.
It is important to remember that this type of model does not have common sense or real knowledge and the computation of correlations between different elements can cause the model to “hallucinate” (yes, this is the technical term) by producing data that is misleading or outright false. A big problem is that these data are produced without the slightest doubt and presented with the appearance of expert opinion. For example, when asking ChatGPT for an explanation for an invented physical phenomenon, the system gave me a completely false explanation (since it is a phenomenon that does not exist) that, however, sounded completely plausible. Another big problem is that all these models contain biases due to the data used for training. Open AI is putting a lot of effort into limiting these problems by using systems to filter this type of content and they have fared much better than previous systems like the infamous Tay, a Microsoft chatbot that “went” racist, misogynistic and anti-Semitic after be learning about social media for just a few hours in 2016.
ChatGPT, and similar technologies, is causing an industry revolution by enabling highly targeted content creation on a scale never seen before. A critical point is the way to ask questions to the system since small changes can have important changes in the result. For example, the command “write a summary of World War II” can be modified by adding “as a history teacher would do” or “for a 10-year-old to understand.” The result is two texts with a similar content but with a completely different style and linguistic complexity.
The WWII example represents one of the most popular use cases. The automatic creation of homework and essays related to education. The impact on education of this technology is comparable to, or even greater than, the creation of the Internet or the use of calculators in the classroom. The education community is very concerned about this technology but, if used properly, it can help regenerate a sector that has not changed significantly for, at least, decades. Other examples include tasks as diverse as generating marketing and promotional articles or advice for all kinds of situations such as the best way to lose weight or how to speak better in public.
One of the darker sides of this technology is that it can be used to produce customised disinformation on an unprecedented scale. In a world with much more content and much more “noise”, the information consumption patterns will rely even more on trust and loyalty to certain media, possibly polarizing us even more politically and socially.
ChatGTP has also been used effectively for tasks with more interaction such as automatic mentors to study new languages. In a more creative case, one organisation used a chatbot to pretend to be a user of a mobile phone network in order to negotiate a price improvement by talking directly, and autonomously, with customer service. Ironically, the customer support system could have been another chatbot itself. Other industries that are being affected include programming or law since ChatGTP is capable of generating both chunks of code and certain contracts as long as it is asked a specific enough question.
In all these cases, the system can (and usually does) produce errors or omissions and its possible impact must be considered. However, effectively combining ChatGPT and humans to solve a task can significantly improve productivity.
The capabilities of this type of technology and its inevitable improvement in the coming years will create a paradigm shift where many professions will use this type of service to avoid doing certain routine tasks and to “not start with a blank page” in content creation tasks. In certain cases, we may even see a change of roles, where humans become an editor and reviewer where our main responsibilities will be to specify the themes and style of the text that must be generated and to verify the veracity of the exposed data.
As in all industrial revolutions, there will be people who will view this technology with suspicion or even fear, but I am sure that we will know how to adapt to improve our lives as we have done so many times before.
The Practical Academic 