[im3w1l]

> Research on judges has yielded similar results. It has found, for example, that external factors as diverse as outdoor temperatures and sports results can influence judges’ decisions.

When I hear of these funky effects I always wonder how they relate to AI. Presumably this is somehow related to some kind of lossy compression of the state of the world, maybe similar to principal components[0]? Where the "I feel grumpy" component is a mix of defendant-is-guilty, temperature-is-low and my-team-lost.

It might also be related to the binding problem[1]?

[0]https://en.wikipedia.org/wiki/Principal_component_analysis

[1]https://en.wikipedia.org/wiki/Binding_problem

[sixhobbits]

Also interesting is "Impossibly Hungry Judges". Many of these correlation effects are paradoxically so strong that they they _cannot_ be the true explanation.

[0] http://nautil.us/blog/impossibly-hungry-judges

[tokai]

"The phenomenon of favorable decisions peaking after a meal break is likely an artifact of the order of case presentation. It is not evidence that meal breaks affect the boards’ decisions."

https://www.pnas.org/content/108/42/E833.full

[mattkrause]

Yup. Even ignoring the lack of randomization, there’s another issue with the hungry judges paper that should have raised some eyebrows.

They fit a model using both the case order (1st, 2nd, 3rd, etc) and the time elapsed. Either is significantly related to the case’s outcome, but when both are included, the rank explains away the time elapsed. This is obviously not compatible with increasing hunger/decreasing blood sugar, since those should depend on wall-time.

[nefoo62]

I don't have a reference handy, but I seem to recall that that study on judges' verdicts being influenced by temperature, how close they were to lunchtime and so on, failed to replicate. I wouldn't be surprised if the OP study fails to replicate as well. They have a relatively high number of data points (to get an idea of order of magnitude: mortality of 145 across 2064 operations on birthdays) but only reach a P-value of 0.03 on their main conclusion.

[mattkrause]

The judges-at-lunch thing turned out not to be a good “natural experiment” because the prisoners were systematically ordered.

The parole board considered cases from one prison at a time. Within each prison, prisoners representing themselves went last and they tended to fair worse than those with attorneys. The judges tended to take meal breaks between prisons, and....poof, there’s the result.

[im3w1l]

This one? https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3645463

[chromanoid]

For human beings: How about simple plain distraction as the problem?

If you refer to AI there are many examples where the training data is biased. One funny example was enemy tank recognition that saw enemies whenever there was gloomy weather, because the sample images of enemy tanks all where shot at such weather conditions to make them appear sinister to human eyes.

If you refer to a mental model, I guess it might simply be a resource management problem. Just because we do not experience the distraction actively it does not mean it is not there. How exactly distraction is compensated is irrelevant to this explanation. Explaining this with mathematical terms is probably pretty arbitrary and leads to framing (in a psychological sense). But I also like speculating on AI ;)