This article first appeared in Il Sole 24Ore, Italy's leading financial daily. TheO provided the translation and minor edits.
Artificial Intelligence, The Response Of Machines And The Role of The University
Santiago Schnell
In 1955, John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon proposed that every aspect of learning and intelligence could be precisely described for machine simulation. They called this research program "artificial intelligence." The term, introduced before the 1956 meeting at Dartmouth College, defined a new field that summer.
Seventy years later, progress has exceeded the original researchers’ expectations. Generative systems now assist with writing, programming, translation, literature analysis, and experimental design. This development prompts universities to reconsider what constitutes a well-formulated answer. As technology advances, we must reevaluate the role of written work as evidence of learning. While essays, solutions, and reports have traditionally signaled student understanding, artificial intelligence has diminished this connection. Rapidly produced texts, however persuasive, may not demonstrate genuine comprehension.
The Aristotelian-scholastic tradition offers helpful terms to distinguish these levels: ratio, the logical progression from one premise to another, and intellectus, the act of grasping a principle and its meaning. Thomas Aquinas viewed these as two modes of the same intellectual activity, not separate abilities. This distinction shows that following an inferential sequence correctly does not guarantee true understanding.
Modern philosophy describes how intelligence became associated with calculation. Hobbes wrote that reasoning "is but calculating," meaning it involves adding and subtracting the consequences of names. In 1950, Alan Turing reframed the question "can machines think?" by introducing the imitation game, which assessed conversational performance without considering the mind's metaphysics. Dartmouth’s proposal adopted a similar approach, seeking to build machines capable of behaviors humans regard as intelligent. While this was scientifically productive, performance alone does not resolve the question of understanding.
The history of linguistic models offers a concrete example. In 1913, Andrei Markov analyzed about twenty thousand letters of Pushkin's Yevgeny Onegin and showed that vowel and consonant sequences were not independent.
Markov measured the statistical structure of text, not the generation of poetry in Pushkin's style. In 1948, Claude Shannon produced English "approximations" by selecting characters based on frequency and prior context. Although these sequences lacked semantic criteria, they still resembled the language. Modern models are more complex than Markov chains, using advanced architectures, large datasets, and sophisticated training. However, many are still pretrained to predict the next token given a context. This approach can yield strong formal competence, but it does not necessarily demonstrate understanding. These facts do not support a metaphysical conclusion.
We cannot determine whether future systems will achieve deeper understanding, and no consensus exists on what evidence would be sufficient. Institutionally, the distinction is clear: qualifications and professional responsibilities are granted to people, not models. When a model generates a hypothesis, diagnosis, or opinion, a person must verify its basis, decide whether to adopt it, and remain accountable for the outcome.
Neuroscientist Stuart Firestein describes science as transforming general ignorance into meaningful questions. Models can review extensive literature, suggest connections, and propose hypotheses, but novelty, plausibility, and truth are distinct concepts. Researchers must decide what to measure and design tests that differentiate between hypotheses. Interpreting ambiguous results and identifying inconsistencies are also essential.
