Marc Teerlink: ‘I, Robot’ scenarios don’t pose a threat despite computers with anticipating capabilities’


IBM BusinessConnect was held yesterday. During this convention, professionals discuss topics such as the way in which Big Data analytics may benefit various business sectors. One of the speakers there was Marc Teerlink, a member of the team that was responsible for the development of the supercomputer Watson. This presented Numrush with a unique opportunity to ask Marc Teerlink a few questions regarding Watson!

Watson is often referred to as a supercomputer. It is particularly famous for beating the big champions of Jeopardy. During the quiz, Watson provided its answers based on what IBM likes to call ‘calculated confidence’. This works as follows: the supercomputer first analyzes the question, and subsequently compares all the sources that are available to it to determine the best possible answer.

In other words, Watson has not been programmed to look for connections, but does so entirely all on its own. In doing so, it can both learn languages and understand their content. This is also called cognitive analytics. Cognitive analytics is an approximation of the feeling we get as people: “what am I more certain of”, or to what extent is one thing we have learned more reliable than the other.


“Reliability is based on different factors. First of all, Watson will ask itself: Have I correctly interpreted the question? This requires context,” explains Teerlink, offering the example of a web search for ‘Paris Hilton’. This does not necessarily mean you are looking for the American socialite; it can also mean you are looking for a hotel in Paris. “In this case, we can respond with another question to clarify things so as to increase reliability,” says Teerlink.

“Second, which sources will I use for my answer? Since I know the context in which the question was put forward, I can say ‘I won’t look in …, but will rather search …’. This has to do with the amount of fact checking the source has done and how often the source has been right or wrong in the past.” According to Teerlink, this is calculated anew with each question, making it possible to remove errors in sources in the database and to factor in a reliability score for the source.

Based on the reliability scores, a list of possible answers is generated. If no answer is reliable enough, Watson will indicate this and say it is unable to answer the question. However, Teerlink keeps stressing that Watson provides recommendations rather than definitive answers.

Watson in the medical world

Watson is now used for pilots in the medical sector. To this end, Watson has analyzed hundreds of thousands of cases and made a diagnosis based on its findings. In many cases this diagnosis turned out to be correct. “Watson is definitely a supporting tool. It offers a second opinion for one’s own opinion as a doctor,” explains Teerlink. Doctors can use Watson to come up with quicker and better diagnoses.

For this purpose, Watson uses case files and patient dossiers. The fact that general practitioners are famous for their illegible handwriting is not an issue, as Watson is capable of machine learning. “The thing about handwriting is that it is consistent. If you go through it once, you will be able to learn from it. For each question, Watson will repeatedly look for an answer. What was the answer last time round? Has something changed since then? No? All the better. If yes, which answer is the correct one?”

I, Robot?

So machines and technology are capable of learning more every time. Cognitive technology is even able to learn how it should deal with people, which means people will no longer have to learn how to handle technology. That said, it looks like we don’t need to be afraid of ‘I, robot’-like scenarios.

“The morality of robots and machines lies with you and me; it is determined by people. We can decide what information we want to share. You can always switch it on or off.” “Every society knows what is desirable behavior. Do we want technology to learn this? I don’t have an answer to this question. What we do see, however, is that a search engine can quickly finish my sentences based on my behavior.” This type of functionality will become more and more common, says Teerlink. There will be an increasing number of sensors that are capable of recognizing language, feeling and movement. More on this can be read on IBM’s Five out of Five website.

“A challenge would be that predictive analytics [making predictions based on data, e.g. regarding personal preferences] gets the upper hand. If so, how can you still discover that one new album? How do you personally discover new things? Therein still lies a challenge for companies and data scientists,” argues Teerlink.

This is something where a role is laid out for marketing, but we will simply have to wait and see how things will develop.