Once a month, Le Devoir d’enseignement wants to offer enriching contributions, whether they come from researchers and practitioners in the education sector or from other people who have reflected on the state of our education system.
Advances in artificial intelligence (AI) are accompanied by a litany of prophecies announcing revolutions in all directions, from manufacturing production to legal practice, including medical diagnosis and artistic creation.
Recently, a television program bringing together the elite of Montreal’s AI community even listed teaching, a field of human interaction par excellence, in the fields that would soon be disrupted by the introduction of generative AI models. This prediction is, however, not surprising since it is part of the chorus of speeches, often alarmist, from self-proclaimed “experts” in pedagogy who announce the inevitable transformation of so-called “traditional” learning and evaluation models. », in anticipation of the massive use of conversational agents by students.
We obviously no longer count the technological revolutions which, over the past 30 years, were supposed to radically transform the world of education. Just think of the interactive whiteboard revolution announced a few years ago, before we realized that, for various reasons, most teachers only used this expensive device as a simple projector.
Distance learning during the pandemic has also revealed the educational limits of virtual school, and even its harmful effects on learning. However, experts are once again affirmative: this time, things will be different with AI.
Taking them at their word, I decided last fall to assign students enrolled in my history of science and technology course a practical exercise which would allow them, in a reflective approach, to evaluate real skills in history of the most popular conversational agent, ChatGPT (in its free version).
The students first had to write a synthetic biography of a relatively well-known scientist or inventor, based on the model of the texts produced in the Dictionary of Canadian Biography. To carry out this work, they would draw on at least three peer-reviewed sources.
Secondly, the students had to produce the same biography, but this time relying solely on interactions with ChatGPT. Thirdly, the students had to compare the two biographies and carry out a critical analysis of the differences observed between the two texts.
One of the objectives of the exercise was to allow students to explore the conversational agent while already having a certain mastery of their subject, in order to more easily identify the limits of “blind” use of this tool. often presented as miraculous.
Wrong facts
First observation, which was expected, the factual errors produced by the algorithm were legion. To a student who had worked on computer science (and AI!) pioneer Alan Turing, ChatGPT claimed that the British mathematician had designed, not deciphered, the machine’s codes Enigma used by the Nazis during the Second World War: a gross error.
He also told her that Turing had carried out a research stay at the Institute of Advanced Studies, in Princeton, between 1945 and 1946, while the student had instead noted the dates of 1936 and 1938 in her own research. Asked about this discrepancy in dates, ChatGPT confessed to having clashed between Turing and one of his contemporaries, the Hungarian mathematician John von Neumann, whose name is rather associated with game theory.
Unfortunately, verification made, von Neumann, although he was in the United States between 1945 and 1946, had not set foot in Princeton. Another student who was working on the inventor Alexander Graham Bell noticed that ChatGPT attributed the authorship of “visible speech” to him, when it was his father, Alexander Melville Bell, who should have been credited with the invention of this system. phonetic.
ChatGPT also taught a student that the French Nobel Prize winner in physics Louis Néel had been trained by Marie Curie before founding his own laboratory at the University of Strasbourg, then taking a professorship at the Sorbonne, as many erroneous statements. In the cases of lesser-known scientific figures, such as the 18th century French astronomere century Nicole-Reine Lepaute, ChatGPT was even more confused by generating entire sections of fictional biographies.
Hallucinations
If these deviations may have made the students smile, they however found it less funny that ChatGPT’s uncontrolled bursts of inventiveness, or “hallucinations”, as AI jargon refers to them by abusive anthropomorphism, extend to bibliographic references.
The student who worked on the physicist Louis Néel had struggled to collect sources to document his work. He was therefore surprised to find that the biography produced by ChatGPT referred to several academic works that he had been unable to find, before being even more surprised to discover that these references were in fact fabricated.
A student who chose to explore the career of doctor Ignace Philippe Semmelweis not only discovered that ChatGPT had suggested non-existent references to her, although they seemed at first glance plausible, but that even the real references he had provided did not mention Semmelweis only anecdotally.
Interesting fact: one of the works mentioned by ChatGPT was even considered a low-quality reference by serious historians of the Austro-Hungarian doctor. Second observation, methodological this time, the conversational agent was not only likely to enrich the historiography with imaginary works, but even when it offered real references, the quality of its literature review could prove to be weak and little relevant.
From an educational point of view, I could have used these invented bibliographic references to explain to students the “mechanics” behind how ChatGPT works. His “hallucinations” are not only due, as is often said, to the fact that the data on which he was trained (roughly, the content of the Internet until 2021) itself contains factual errors or contradictory and biased information, since the erroneous references it produces simply do not exist on the Internet.
These “hallucinations” are in reality inseparable from the tool itself, which remains a very powerful generator of… probabilistic texts, forming sentences from the probability that words appear in similar sentences and contexts. In other words, neither intelligent nor creative, ChatGPT is an algorithm that relies on statistical methods of calculating probabilities and a massive amount of training data to generate the text with the highest chances of responding “correctly” to a question asked of him.
Even if it were trained on a corpus of “perfect” data, the probability of it generating errors would not be zero. ChatGPT therefore responds in probabilistic terms and not according to truth criteria; its “intelligence” is therefore only apparent, as is that of all algorithms.
A simplistic approach
The third limitation identified by some students in ChatGPT’s prose refers to the very nature of what a good scientific biography should be. Several students noticed that the texts generated by the conversational robot regularly veered into hagiography and presented scientists or inventors as individual geniuses and solitary heroes of science, thereby obliterating the social and intellectual context that had influenced their path.
A student who took Antoine Lavoisier as a subject noted that, contrary to his text, that of ChatGPT had failed to situate the French chemist’s discoveries on oxygen in relation to the experiments of his British contemporary Joseph Priestley, who relied on the concept of phlogiston. This context is, however, crucial for understanding the originality of Lavoisier’s scientific approach and the epistemological break that it makes with the qualitative approach that prevailed until then in chemistry.
Even more blatant is the case of Thomas Edison, presented by ChatGPT as the “inventor” of the electric light bulb, an assertion which is common sense but which is devoid of historical depth. Indeed, the development of the incandescent lamp with carbon filament resulted rather from a collective work of invention, Edison himself being at the head of a team of inventors employed in his laboratory at the end of the 1870s .
At the time, Edison was far from the only one working on an incandescent lamp concept, as evidenced by his association with the British electrician Joseph Swan. The success of his model was also dependent on the use of other innovations, primarily the mercury pump developed by the German-British chemist Hermann Sprengel in 1865.
To present Edison solely from the angle of the inventor endowed with natural genius is to forget that what made his social existence possible was the emergence of industrial research, which began to be organized into collective activity in the companies at the end of the 19th centurye century.
Although ChatGPT generated impeccable-looking biographies, composed of well-structured sentences and free of grammatical or syntax errors, it also showed significant limitations both in terms of the precision of the facts presented and the relevance of the sources provided only from the point of view of the problematization of his biographical subjects.
For teachers, this undoubtedly makes it a very good reflective tool to use in historical research methodology courses. For students, it is a tool that can undoubtedly prove useful in simplifying certain tasks (translating extracts of text, suggesting work topics or avenues for reflection, improving the quality of their writing).
Nevertheless, the exercise carried out by my students shows that before conducting an in-depth exchange with ChatGPT on a given historical subject, a reasonable knowledge of the subject in question remains an essential prerequisite in order not to be fooled by its multiple traps.
*With the participation of students from course HST1143.
Suggestions ? Write to Dave Noël: [email protected]