Abstract says "mean absolute error of 0.229 °C"

It’d be hard to get average error so low, since you need precision to get low error values for varying temps.

You can’t just return 100 and -100 randomly, since the actual temp is unknown and you’d need to evenly sample around that.

A normal distribution will average to 0, with enough samples. The center and spread is important, since individual measurements can be considered a random sampling within that distribution. But, this is overly simplifying things, and assume random error, which AI systems often don't exhibit.

It appears this is moot, because the quote did not reflect the actual value or method used.

The distribution of human temperature levels is already known, you could just give the average temp as every guess and get maybe 1.5 F average error.

My point is, you can’t really get 0.44F avg error unless it’s predictive. The “std dev” of abs(reality - model) is too high

I would have naively calculated average error as avg(max( abs(smallest), largest)). so it would be 100 in your case. I'm sure that's not statistically correct but it makes sense to me.