The A-Z of AI

Watson

The first AI system to beat a human contestant on a TV game show.

In 2011 IBM created Watson — an AI system that challenged two humans to a game of the U.S. quiz show "Jeopardy!" and won.

Watson based its answers on the contents of millions of books, dictionaries and encyclopedias.

Once Watson had been taught the unique structure of Jeopardy! questions from previous examples, the system developed a way to find the answers to new questions within a fraction of a second.

Programmers trained the AI to not only search for keywords within content but to also analyze and decipher the sentences around those keywords. That way, Watson could cross-reference different contextual sources to work out that "tea" was not only a drink, but also a type of dress from the 1940s (and the answer to an $800 question).

However, if the host asked a question in an unexpected way, using a sentence structure that the programmers hadn’t anticipated, Watson would fumble. Watson’s lack of conversational context, which humans naturally have, also made some questions almost indecipherable to the AI.

This proved instrumental to the scientific community, as it shined a light on the need for flexible AI systems trained on robust examples of human speech and informed the way that natural language systems are built today, such as those that power the virtual assistants we have conversations with in our homes and on our phones.