Professor Benoît Garbinato shares his insights into the mechanisms of generative artificial intelligence. He explains how these systems operate based on probabilities, highlights the crucial role of human input in their training, and explores the ethical and educational implications of their use. His message is clear: these tools are powerful, but they must be understood in order to be used wisely — especially in higher education, where they are already challenging traditional assessment methods.
Professor Benoît Garbinato from department of Information Systems at UNIL Business faculties (HEC), shares his insights into how generative AI works and its implications for teaching. With an engineering degree from EPFL, Professor Garbinato began his career in industry before transitionning to academia in 2004. Although artificial intelligence is not his area of specialization, his strong background in algorithms enables him to offer an informed perspective on these technologies.
A system based on probabilities
Professor Garbinato explains that the working of a langage model as ChatGPT is based on a similar principle as text prediction models on smartphones: d’un modèle de langage comme ChatGPT repose sur un principe similaire à celui des systèmes de prédiction de texte des smartphones : “It’s very simplified, but it’s the same kind of algorithm you have on your phone when you’re trying to write a message and the software suggests the end of your sentences.” The key distinction is the scale: ChatGPT has been trained on an extensive corpus and operates with millions of parameters.
A crucial point raised by the researcher is that these models have no real understanding of what they generate : “ChatGPT hasn’t any notion of what is true or false, what is logical or illogical. However, what it has is a notion of what makes a statement plausible and fluent.”
Human intervention in data cleaning
The interview highlights a lesser-known dimension of generative AI development: the substantial human effort involved in cleaning training data. In order to mitigate the reproduction of biases and harmful content present on the internet, a large workforce is tasked with curating and filtering the textual datasets used during model training. This situation brings to light a number of ethical concerns, especially in relation to the labor conditions of the individuals performing this work—many of whom are underpaid and based in developing regions.
As an illustrative example, the professor mentions Microsoft’s AI Tay in 2017-2018, which had to be shut down after it became “racist, misogynistic, and conspiracy-minded” in just 24h of exposure to Twitter contents. This experience then highlighted the importance of pre-filtering training datasets to prevent undesirable model behavior.
Implications for education
Professor Garbinato, speaking as an educator, describes these tools as ‘absolutely fantastic—provided you know what you’re doing with them.’ He also highlights the importance of making a qualitative evaluation of the sentence generated. This technological shift is prompting him to rethink his assessment methods: ‘Testing rote memorization no longer makes sense. On the other hand, understanding how a piece of code works, identifying its issues or vulnerabilities—that requires real intelligence.’
He encourages his colleagues in education to prioritize conceptual understanding over rote memorization, given that AI tools are now capable of producing highly plausible discourse without any genuine comprehension.
Developpment prospects
Regarding the future of these technologies, the professor notes that the initial wave of enthusiasm is beginning to subside. He points out that ChatGPT’s true innovation was less technological than social—stemming from its release to the general public.
Looking ahead, he mentions ongoing research—particularly at Google—aimed at combining language models with a ‘logical truth model.’ ‘The day we manage to integrate the two, we’ll take a significant step forward. […] Not only will we be able to generate fluent text, as we already can, but we’ll also have some guarantees that the output is logically coherent.’
In conclusion, Professor Garbinato emphasizes the importance of not treating technology as magic. He urges users to always verify the information generated by these systems, even when it appears plausible: ‘Even if the discourse is fluent, even if it seems logically well-structured, go check the facts, cross-check the information, and see whether multiple sources say the same thing—before concluding that because it sounds fluent and credible, it must be true.’