Self-assessment in expert AI predictions

This brief post is written on behalf of Kaj Sotala, due to deadline issues.

The results of our prior analysis suggested that there was little difference between experts and non-experts in terms of predictive accuracy. There were suggestions, though, that predictions published by self-selected experts would be different from those elicited from less selected groups, e.g. surveys at conferences.

We have no real data to confirm this, but a single datapoint suggests the idea might be worth taking seriously. Michie conducted an opinion poll of experts working in or around AI in 1973. The various experts predicted adult-level human AI in:

  • 5 years: 0 experts

  • 10 years: 1 expert

  • 20 years: 16 experts

  • 50 years: 20 experts

  • More than 50 years: 26 experts

On a quick visual inspection, these results look quite different from the distribution in the rest of the database giving a much more pessimistic prediction than the more self-selected experts:


But that could be an artifact from the way that the graph on page 12 breaks the predictions down to 5 year intervals while Michie breaks them down into intervals of 10, 20, 50, and 50+ years. Yet there seems to remain a clear difference once we group the predictions in a similar way [1]:

This provides some support for the argument that “the mainstream of expert opinion is reliably more pessimistic than the self-selected predictions that we keep hearing about”.

[1] Assigning each prediction to the closest category, so predictions of <7½ get assigned to 5, 7½<=X<15 get assigned to 10, 15<=X<35 get assigned to 20, 35<=X<50 get assigned to 50, and 50< get assigned to over fifty.