Albert J. Menkveld

My two cents

AI "Errors"

As researchers increasingly seem to use AI to do their analyses, one wonders to what extent AI outcomes differ from the pre-AI era of human outcomes. One can only find out if there somehow is a “gold standard” of human outcomes to compare to. This is where the #fincap experiment is extremely useful. It had 164 human teams independently estimate six annual trends on the same sample in a pre-AI era.

In a new paper with Wenqian Huang and Shihao Yu, we give AIs the same set of instructions as the humans got in #fincap. The two main questions we ask are:

1. Do the outcome distributions differ when comparing AIs to humans? We refer to differences as “AI errors,” considering humans as the gold standard (which can be debated).
2. If they do differ, can we somehow find the decision forks on the analysis path that are responsible for the largest differences?

The box plot summarizes, for each of the six trends, the distribution of percentage-annual-change estimates.¹ The blue box plots are based on 164 estimates of the human teams. The red box plots correspond to 158 AI estimates. The box stretches from the first to the third quartile, with the horizontal line indicating the median. The whiskers extend to the 2.5% and 97.5% quantile.

The plot, along with an analysis of key forks that drive a wedge between the two distributions, led to the following insights:

• Both outcome distributions exhibit substantial variation. It seems that neither humans nor AIs seem to agree internally on the “right” outcome.
• We find that the distributions are statistically different.
• The AI distribution often showing less variation. An important explanation for this lack of variation is that AIs show less variation across all possible analyis paths. They tend to concentrate on fewer options at the fork when compared to humans.
• A particularly important decision fork is the choice of statistical model. An interesting observation is that AIs seem to avoid the trap of estimating a trend by computing the average relative change. This can create a substantial negative bias in estimates, in particular when the data is sampled at high frequencies to compute a low-frequency trend. (The argument relies on Jensen’s inequality, which is a bit too technical for this post. It is in the paper.)

This final point is visible in the box plot distribution where the negative tail can reach out to exceed -100,000% for humans. This is not happening for AIs. One could say that this is a benevolent type of AI error. Yes, it is the source of a wedge between AI and human distributions, but it undoes a bias.

“You can tell it’s AI. No left tail.”

Please find the paper here.