Flash Fill and Series Fill are great at recognizing simple patterns like 1,2,3,4,5 or 2,4,6,8,10 and can sometimes (unreliably!) recognize more complicated multiplications, repetitions or formulaic string modifications (like house numbers and street names). I use both frequently.

But neither these tools nor an LLM has real understanding, they're not going out into the real world and collecting data. The results need to be verified. Automatically filling an index column with 1,2,3... is one thing, automatically filling a data column with guesses from pattern matching is different. This problem is only going to get worse as LLMs proliferate and can more reliably but still imperfectly and still opaquely fill in more complicated data.

I don't use autofill precisely because I have no insight into the algorithm behind it. I suspect it's very "stupid," but I don't know exactly how stupid it is.

It sounds like he didn't realize or notice that autofill was using adjacent cells (from other countries, in this instance). That behavior would shock me as well. I thought autofill only took columns of data into account.

> It sounds like he didn't realize or notice that autofill was using adjacent cells (from other countries, in this instance).

No, it doesn’t sound like that: “in several instances where there were no observations to use for the autofill operation, the professor had taken the values from an adjacent country in the spreadsheet.”

It was not Excel’s fault.