ChatGPT: A Technological Breakthrough?

Capable of conversing with a human, structuring a response, and rendering a certain amount of knowledge, the ChatGPT can also write and debug code, compose music, play games, answer evaluation questionnaires, write poems and songs (and more).

According to Laure Soulier, an associate professor in computer science in the Machine Learning for Information Access Team at ISIR¹, this new tool is in line with the large language models developed in recent years, which are capable of processing a very large amount of data and solving increasingly varied and complex tasks. In a few months," she says, "we have seen the emergence of many interesting language models, more or less specialized, such as Galactica or LaMDA." For the researcher, ChatGPT is not a breakthrough, but rather an improvement of existing models and contributes to democratize artificial intelligence to the general public through an accessible conversational interface.

To reach this level of efficiency, OpenAI engineers used the GPT 3.5 algorithm trained on numerous documents (Wikipedia, web articles, forums, etc.) and the instructGPT model, which refines the training of language models by integrating human judgments. Engineers also took and adapted the instructGPT data to direct it towards dialogue tasks and obtain better relevance in the answers.

"The training of ChatGPT consists of several steps. In the first stage, two people are asked to converse to obtain a data set with one simulating the system and the other the user," explains Laure Soulier. We learn a first language model dedicated to the conversation. We give an example to the machine, we see what it responds to and, depending on its level of error, we modify the parameters until it reduces it. Once the model has learned from this data, the second part of the training consists in asking it to generate several outputs for the same conversation. Annotators will then rank the model's responses in order of relevance. This allows us to have a second supervised dataset and to learn a reward model," says the researcher.

The final step is to refine the model learned in the first step by augmenting it through reinforcement learning². The latter uses the reward model learned in the second step to readjust the parameters of the model. Once trained, the model can be used to generate text autonomously.

Ethical and technological limits

Fast and accurate, with results that seem almost magical, ChatGPT raises its own set of ethical questions, including the credibility and veracity of the information. "It writes plausible answers that seem consistent, but may in fact be inaccurate or misleading. Training based on reinforcement learning does not force the model to generate truthful information and its knowledge of the world is limited to the data it has been given during training. As it is not currently connected to the web, the ChatGPT does not have access to new information published. Other biases exist, adds the computer scientist, such as the tendency for language models to over-generate. Moreover, this model is opaque. We don't know the data on which it was trained, nor how the annotation was done, nor the instructions given to the human annotators.” Another question is whether ChatGPT should answer all questions. Open AI has already evolved its model so that it no longer provides the recipe for creating bombs on demand.

The question of plagiarism and copyright is also at the heart of the debate. Who owns the generated text? The user who asks the question? In an attempt to combat potential plagiarism, Open AI is currently developing a system to automatically detect chat-generated text. "We had the same concerns when Wikipedia came out. But in the end, Wikipedia is a medium that will never replace school. In the same way, ChatGPT will allow us to have an overview of many things that we will then have to rework. This raises the question of the user's place," says the researcher. Unlike ChatGPT, a search engine allows the user to access sources, to make sense of the information he finds and to be active in the knowledge construction process. "Conversational systems don't necessarily allow for this diversity by providing direct answers. This is an area of information retrieval that is currently in vogue. One of the theses that I am directing, deals precisely with the clarification of questions in order to benefit from the advantages of search engines and conversational systems," explains Laure Soulier.

Faced with these challenges, the role of the scientific community that has been working for years on these language models is great. "Everyone is responsible for how they produce the models, make them available and use them. Once they exist, they cannot be ignored. The question now is what are the next steps and how can we use these advances to improve other models and other tasks?" she says. Other models with knowledge of the world that is not limited to textual data, but to all data (visual, audio, etc.) are being developed.

According to Laure Soulier, the scientific community also has a role to play in raising awareness and educating people about the design and use of these models and their limitations. "Even if these tools are not intended to replace the user, but to help him or her become more efficient, it is important that people develop a critical mindset towards them and the biases they entail," she concludes.

¹ Institute of Intelligent Systems and Robotics (Sorbonne Université/CNRS/Inserm)

² Reinforcement learning consists of an autonomous learning system from experiments with the environment by observing the positive or negative results of its actions. It allows machines to determine automatically the ideal behavior to adopt in a specific context in order to maximize its performance.