By: Elon Danziger, freelance art historian / The New York Times
Translation: Telegrafi.com
Opposite Florence Cathedral in Italy stands a much older church, the Baptistery of St. John. It is a beloved center of religious life, where many Florentines are baptized to this day. Simple columns and ornate arches hug its eight sides, half-camouflaged in green and white marble patterns. Without the baptistery’s imitation of ancient Roman architecture, it is hard to imagine Florence giving birth to the architectural Renaissance that changed the face of Europe. Yet for centuries, there has been no convincing answer to the questions of who built it, when, and for what reason. Decades ago, I would organize visits to the baptistery and worship it, and in the early 2020s, I began to delve deeper into its origins.
After years of poring over historical documents and reading voraciously, I made a significant discovery that was published last year: the baptistery was not built by Florentines, but for Florentines—specifically, as part of a collaborative effort led by Pope Gregory VII, following his election in 1073. This discovery occurred just before the explosion of artificial intelligence [AI] into the public consciousness, and I recently began to wonder: Could a large language model, like ChatGPT, this puzzle faster than me?
So, as part of a personal experiment, I tried to pass three artificial intelligence chatbots - ChatGPT, Claude and Gemini - through various aspects of my investigation. I wanted to see if they could recognize the same traces I had found, assess their significance, and reach the same conclusions I had reached - eventually. But the chatbots failed. Although they were able to process dense texts for important information about the origins of the baptistery, they were ultimately unable to construct a completely new idea. They lacked the essential qualities to make discoveries.
There are several reasons for this. Great language models have read more text than a human could ever hope to. But when AI reads text, it simply selects the models. It can miss the unique details, the extraordinary insights, and the unusual perspectives that can influence the way we think. For example, in his 2006 book, Tuscany Romanesque, Guido Tigler - a professor at the University of Florence - argued that the baptistery was built later than generally believed. It is an idea that is not widely accepted, and I believe that is why the chetbots never presented it to me when I asked them what they would read to solve the mystery of the baptistery. Although I eventually found reasons to reject that later dating, Mr. Tigler's unorthodox ideas taught me to take more seriously the possibility that previous studies could be wrong about the chronology of the baptistery.
For centuries, many people believed that Pope Nicholas II consecrated the baptistery in 1059. In fact, there is no known evidence of such an event; its existence is based on an assumption drawn from several documents that show its inclusion, in that year, in other Florentine churches. When I pushed the chetbots to discover this discrepancy for themselves, ChatGPT and Claude they found it, but failed to notice that it was suspicious, while Gemini created false evidence that would eliminate this discrepancy. To contribute to a field of knowledge, you have to survey the terrain accurately, sniff out what is suspect, and show why it is wrong. Large linguistic models struggle in all three of these respects.
And here’s the deeper problem: sometimes pattern recognition—both by humans and machines—is wrong. Even though there was no confirming evidence, most scholars had simply assumed that the baptistery’s financiers were Florentine. After all, the vast majority of church building in the Middle Ages was directed by local people: bishops, abbots, wealthy families. But from my reading, I increasingly came to agree with a fringe view that the people of 11th-century Florence were still too poor and provincial to have produced such an accomplished structure.
The key to identifying who built the baptistery was how much its architecture was inspired by the ancient Pantheon in Rome. By the 11th century, the Pantheon had become a church and belonged exclusively to the pope. Once you remove Pope Nicholas from the equation and focus on the popes obsessed with ancient Rome, only one name comes to mind for the mysterious patron: Gregory VII.
A few years before Gregory’s election in 1073, the Florentines had banned infant baptism in Florence, fearing that a corrupt bishop would be unable to protect the souls of their infants. After an event that proved the bishop’s unworthiness and led to his expulsion, the powerful rulers of Florence (and all of Tuscany), Beatrice of Bari and her daughter Matilda, seem to have improved the city by working with Gregory to give it a magnificent baptistery. The rich splendor of Romanesque architecture in the heart of Florence is precisely the kind of church that Gregory would have supported.
Synthesizing so many pieces of medieval history into a new interpretation required stepping back and reexamining their significance and how they connected to each other. AI might be able to match the process of piecing those pieces together, but discovery means making new connections—something that is beyond the current capabilities of artificial intelligence, as my tests confirmed.
Discovery remains a human endeavor and is driven by that deeply human quality - to see unusual things that don't fit the patterns and to investigate them more deeply. /Telegraph/
