August 15, 2012

The end of the Maes-Garreau Law?

AI probability density (skew gaussian and triangular)How much should we trust predictions about future technologies? There are some technologies that are always a few decades away (fusion is the canonical example), there are others that seem to strike without any prediction (stem cells? the WWW?) and others that seem to progress reliably so far (Moore's law).

Stuart Armstrong and Kaj Sotala has produced an excellent post analyzing a set of predictions about the future of AI. Among other things, they looked for evidence of the Maes-Garreau law, that people predict AI somewhere about when they retire. Somewhat surprisingly, they found that this was not true. Instead, over a third of predictors claim AI will happen 16-25 years in the future, irrespective of age. There was no strong correlation between age and expected distance into the future.

This is not necessarily good news for the predictions of AI, since they also found that there is little difference between experts and non-experts, and there is little difference between current predictions, and those known to have been wrong previously. While this doesn't prove that people don't know what they are talking about, it should give us a certain skepticism against confident claims.

It is also notable that the "law" has been relatively widely cited, but the underlying data supporting it seems to be a handful of claims. Kevin Kelly gives a list of 17, some of which are apparently erroneously cited. I think we should add another law of human nature to the laws discussed in the New Scientist article: as soon as a claim has "law" added to it, people will start taking it far more seriously than it should.

There is a bit of irony that a law intended to make people more cautious about confident claims about the future is actually so rickety itself.

Posted by Anders3 at August 15, 2012 03:51 PM