51¶¯Âþ

Artificial intelligence researchers are not publishing enough failures, unlike their colleagues in mature fields such as physics or medicine. This is a shame, say Ehud Reiter and colleagues at the University of Aberdeen in Scotland, because well-documented failures help researchers refine their theories. To set a good example they have confessed to their own recent failure. In the journal Artificial Intelligence (vol 144, p 41) they describe a computer program designed to encourage people to quit smoking. It exploits previous research showing that people are more likely to give up if encouragement is tailored to them personally. Reiter’s team decided to try to automate the process by creating a computer program that examines the responses of 2553 participants to a four-page questionnaire about their smoking habits. They then used an AI technique called natural language generation to produce a personalised letter to encourage each of them to quit. It didn’t work.