Indovinello Veronese

Welcome to another podcast of Between the Lines.  Here we explore the limits of AI and the hidden potential of human intelligence.  Today, we travel back in time…

It was the early 9th century.  And one day, a medieval scribe took up his quill and began to write.  He left no trace of his identity.  He did not sign his name.  All he left was a riddle.  But who or what is he talking about?   So here’s what he wrote:  

He led oxen in front of him.  

He ploughed a white field.  

He held a white plow.  

He sowed a black seed.

Now, let me slowly read the riddle again and take it apart line by line.

He led oxen in front of him.  

He ploughed a white field.  

He held a white plow.  

He sowed a black seed.

Once the metaphors are figured out, the riddle reveals its secrets.  The subject of the riddle is the writer himself.  The oxen are his fingers which draw a feather (the white plow), across a page (the white field), leaving a trail of ink (the black seed).  

The riddle, now called the Indovinello Veronese, was discovered in 1924 by Luigi Schiapparelli.  It is one of the earliest written examples of vernacular Italian.  

We humans can quickly put two and two together.  That is, we immediately “see” the riddle’s solution, once the meaning of the words on the vellum are linked to the physical act of writing.  From there, our imagination takes over, opening up new vistas, delivering new insights.  This is the essence of human intelligence.  It’s creative.  Imaginative.  It reads between the lines.

But can a computer read between the lines?  We decided to see if ChatGPT could solve this riddle and arrive at an interpretation equivalent to a human response.

We began the ChatGPT session by telling it that this conversation would focus on a riddle.   The model responded by saying it “loves” riddles and would do its best to solve it.  We then provided the full-text of the Indovinello Veronese.  Within seconds, ChatGPT responded that this riddle is about a farmer, and it offered commentary for each line of text.

At a superficial level, this response is correct.  The riddle speaks of plows, fields, and sowing – all words one might associate with a farmer.  Clearly, ChatGPT had done its homework, interrogated its word embeddings, and discovered a set of relationships between farm related words. 

The author of this riddle, on the other hand, was not constrained by a limited set of word embeddings but only by his imagination.  The creative leap he took was to compare the act of writing to the work of a farmer in a field.  Although the tools are different, our brains quickly see the connection, once it's made.    

Apparently, ChatGPT did not execute a search to find this riddle on the internet.  If it had, the model would have quickly found its Wikipedia entry, with the solution clearly stated.  The Indovinello Veronese is about the writer himself. 

Pressing ChatGPT further, I asked, “Are there any other possible answers?”  The machine replied  – death and time.  Though it tried to explain its reasoning, the model had largely moved beyond the norms which govern symbolic or metaphoric language.  The answers simply didn’t make much sense.  Even so, this raises an intriguing question: how does the human mind know where the limits of symbol and metaphor lie – of what is acceptable and unacceptable linkages?  Is this an innate ability?  Or something we learn while playing the language game? 

Special Acknowledgment: I want to thank Dr. Mary Watt at the University of Florida for sharing the Indovinello Veronese with me and suggesting it as a possible test case.