ChatGPT: What is it and what’s next?

Martha LealData Protection Lawyer, Master in Law and International Business at Universidad Internacional Iberoamericana Europea del Atlántico and Universidad Unini México, ECPB Data Protection Officer at Maastricht University, certified as Data Protection Officer by Exin and FGV-RJ and president of the Commission for institutional communication of the INPD

Pedro Fratini ChemMaster in Artificial Intelligence and Data Science at the Université Grenoble Alpes

Chances are you’ve heard of ChatGPT before. And perhaps you’ve also interacted with technology and been amazed at your ability to understand the questions asked and the appropriateness of your answers.

However, regardless of your level of knowledge, it’s important to make some basic considerations about the topic and its likely impacts for a better understanding of the problem.

We have chosen to start by deciphering what this technology is about and what makes it different from the others, thus allowing us to reflect on its own evolution and foresee a future in which we will inevitably be increasingly connected to the advancement of artificial intelligence.

ChatGPT is a language model that can predict responses in a chat environment. It has been adapted through an existing model, the GPT-3.5, to behave like a chatbot, interacting with a user in a chat environment where the responses provided by the model appear to come from a human.

For a better understanding, it is important to understand the meaning of the acronym GPT; “Generative Pretrained Transformer”. It is an artificial intelligence model with general purpose linguistic understanding capabilities, i.e. it is not a model focused only on a single task, such as translating texts.

This AI model is characterized as a “deep learning” process, called deep learning, that uses neurons interconnected in a layered structure, similar to the human brain.

OpenAI, the AI ​​research lab that created ChatGPT, unveiled the first version of the technology in 2018, but it wasn’t until 2020 that the technology gained attention with its third version, GPT-3, which it had a model trained with 175 billion parameters and a much more robust text base.

Now we are in the fifth version and the technology has improved techniques that allow you to receive more complex instructions and, consequently, to generate much more coherent responses.

Add to this the fact that the GPT-3.5 version was trained with an immense base of texts available on the web, including posts on social networks and content available in the media.

ChatGPT learns from user feedback, making its responses more and more coherent, which allows us to have good expectations about the constant improvement of technology based on user interactions.

For the purpose of a better understanding of the functioning and potential of the tool, we have selected the answer provided by ChatGPT itself to one of our interactions, when we asked ourselves about its functioning.

Here is the answer: “ChatGPT is a language model that has been trained to generate responses to user questions and commands in a way that makes it appear as if you are conversing with a human. It was built on a large amount of text data and is been trained using machine learning techniques to learn how to produce appropriate and consistent responses to different types of input. In summary, ChatGPT is a language tool designed to simulate a human conversation in a natural and consistent way. It can be used in various contexts to provide responses automation and interactions with users.”

The above example illustrates that a simple user command for the technology to briefly describe its behavior already grants its ability to appear human, providing information and advice to what is asked of it. And there is no doubt that this is just the beginning of a disruptive way of using AI, where when we seek information in the form of questions and answers, on the most diverse topics, we are provided with knowledge, references and, depending on the interaction formulation, with counseling.

Thus was born a new form of search engine, efficient and friendly in terms of user experience. However, like any language model powered by a large volume of data, ChatGPT has risks and impacts associated with its use, such as bias and discrimination, among others.

Among some technological risks that deserve to be highlighted, we have:

The tool is capable of serving a great educational purpose, aiding in problem solving and new learning, but it does require that it be used responsibly. Although some results are impressive for their degree of coherence and robustness, this does not mean that the information provided is always true or that it doesn’t even make sense, leaving it up to the user to verify the veracity of any information presented by the linguistic model, even if the language used demonstrates full conviction and ownership of the matter in question.

Another relevant point is the ability of ChatGPT to become an excellent search engine. There is no doubt that, with all its potential, it is able to compete or support tools on the web, such as Google. In fact, there are already rumors that the model could be incorporated into Microsoft’s Bing tool.

With the current scenario, it is possible to hypothesize that it is likely that linguistic models like these currently being created and improved, will not replace, but will complement search engines, due to their different purposes. Although ChatGPT offers a better answer to users’ questions, it is still not possible to use it to browse the web.

In the field of Health, a study recently published by Drexel University (USA), has demonstrated that ChatGPT-3 is able to provide aids that indicate the onset of dementia with an accuracy of 80%, based on the user’s natural language that interacts with the tool. In the words of the author of the research published in “PLOS Digital Health”, Felix Agbavor, the speech analysis that makes the AI ​​model in question, makes it a promising candidate for recognizing subtle changes in speech that could indicate the onset of an image of madness.

It is therefore imperative to recognize the potential of technology, its benefits and the risks arising from its misuse. In the pursuit of ethical and responsible AI, it is essential to ensure that the technological training base is diversified and balanced, reducing the risk of reproducing prejudice and discrimination.

Starting from the recognition that ChatGPT is trained from a large database, the logical corollary is the conclusion that, if by any chance, this base is contaminated by discriminatory bias, these patterns will consequently be reflected in the responses generated by the model.

In this sense, we recognize the greatness of the technology, the incalculable potential of its benefits, as well as the challenges imposed, natural for all disruptive models, and which confirms the importance of developing and strengthening artificial intelligence regulations, such as the Proposal of regulation of Artificial Intelligence in the European Union, IA ACT and Proposal for regulation of Artificial Intelligence in Brazil, presented at the end of the year to the Federal Senate.