Turing Test Tournaments

Attempts to mimic intelligence in machines have made big strides by taking advantage of "big data".

Essentially, it doesn't require as much genuine intelligence to produce correct behaviour - if you have a big cache of previous correct solutions to draw on.

While the Turing test is an inelegant way to measure intelligence in machines, it has historical significance as one of the first tests proposed.

The first solutions to the Turing test will probably take advantage of numerous existing Turing test transcripts. There are existing Turing test transcript repositories - for example, The Loebner Prize in Artificial Intelligence makes its transcripts available online. However, more data would be useful and welcome.

This page is concerned with a proposed way to create a large Turing test transcript repository - by holding online tournaments in which both humans and computers interact and compete.

Both humans and computers would compete. Two skills could be tested: the ability to act like a human, and the ability to ask questions so as to rapidly distinguish between human and machine participants. Machine entrants could compete in one or both roles.

Online Turing tests face the issue of how to authenticate participants. This becomes especially significant if there are entrance fees and prizes. No authentication is the simplest approach. In practice this means that rewards need to be confined to not dropping out of the contest before being graded - and acting convincingly like a human.

While it would be nice to reward correct identifications of other participants (as humans or computers) and penalise incorrect identifications, this approach seems incompatible with economic rewards - because of the risk of encouraging cheating.

There have been previous online turing test tournaments - e.g.: http://www.turinghub.com/.

Tim Tyler | Contact | http://matchingpennies.com/