True, plausible-indeed, likely. AI as an opaque, statistically reliable oracle

About this session

Artificial intelligence does not tell the truth. Not even what is merely plausible, for that matter. It tells what is probable—more precisely, what is statistically least wrong among a billion possible attempts. And yet we treat it like an oracle: instant answers, confident tone, no hesitation. But beneath the glossy surface of generative and predictive algorithms lies a troubling opacity: no one—not even those who designed them—can truly explain why the AI responded the way it did. Its certainty is only statistical: it works on average, not in the individual case. This talk dismantles the illusion of the transparent oracle and exposes the radical difference between being true and simply not having been disproven by the distribution curve. Between truth and plausibility, a void opens up—one that AI fills with probabilities trained on past data, often biased, always finite. The result is a paradox: the more statistically reliable AI becomes, the greater the risk that we mistake the probable for the authentic. A provocation to rethink our relationship with machines that do not know what truth is, yet are remarkably good at pretending they do.

Speakers

Nicola Donti

Docente Filosofia del Linguaggio e Filosofia della Scienza